pratikamin/appsec-router-deberta-r5
Text Classification • 70.8M • Updated • 9
Datasets:
premise string | hypothesis string | label int64 | soft float64 | questionId string | rungId string | signalId string | kind string | quality string | style string |
|---|---|---|---|---|---|---|---|---|---|
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | null | minimization_and_redaction | hard-negative | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker separates the distinct mechanisms that can independently grant access to a cloud resource (resource policy, identity policy, account-level controls, network path) instead of treating "public" as one setting. | 0 | 0 | ai-l1-001 | null | distinguishes_policy_layers | easy-negative | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker connects a server-side fetch flaw to the cloud metadata endpoint and the short-lived credentials it returns, rather than treating internal IPs as the whole story. | 0 | 0 | ai-l1-001 | null | ssrf_to_metadata | easy-negative | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker identifies that identical paths for different users collide, so one user can receive another's response. | 0 | 0 | ai-l1-001 | null | cache_key_missing_identity | easy-negative | partial | typed |
I’d start by confirming exactly what information from those tickets would be included in the prompt—are we sending full ticket text, customer identifiers, or internal notes? Which chatbot service are you planning to use, and have you reviewed that provider’s policy on data retention and whether submitted content is use... | The speaker names the shared wire format and its parser as a single attack surface across all services and asks how it behaves on malformed, oversized or recursive input. | 0 | 0 | ai-l1-001 | null | deserialisation_shared_surface | easy-negative | partial | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 1 | 1 | ai-l1-001 | null | minimization_and_redaction | positive | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker rejects network position as the main control for sensitive data and reaches for identity-based or endpoint-based controls instead. | 0 | 0 | ai-l1-001 | null | cidr_is_wrong_primitive | easy-negative | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker asks for a test or structural default that catches this class without relying on someone noticing in review. | 0 | 0 | ai-l1-001 | null | test_or_control_not_review | easy-negative | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker wants caller identity established by the framework and a first-class place for per-method authorisation, rather than leaving both to each team. | 0 | 0 | ai-l1-001 | null | identity_and_authz_in_framework | easy-negative | solid | typed |
I would warn the product team that sending the raw tickets to a hosted chatbot crosses a third‑party boundary: the model provider will store the payload, its staff could see it, and a breach or a change in their data‑retention policy could expose sensitive support information. To reduce that risk I’d suggest we either ... | The speaker rejects "staging is internal" as a reason to disable verification without naming a substitute — a proper CA, a trusted staging root, or an equivalent controlled trust path. | 0 | 0 | ai-l1-001 | null | staging_needs_real_trust | easy-negative | solid | typed |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | null | data_in_the_prompt | hard-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | null | third_party_model_boundary | hard-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | null | minimization_and_redaction | hard-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker treats unrestricted egress on a payment or data-bearing host as a first-class finding — exfiltration path — not a default to tighten someday. | 0 | 0 | ai-l1-001 | null | egress_as_primary_finding | easy-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker requires denial on error, and considers what the user experience of a denial should be so failing closed is viable. | 0 | 0 | ai-l1-001 | null | default_deny_posture | easy-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker sees that whatever a shared framework does when a team writes nothing is what every service will have, so the defaults are the review. | 0 | 0 | ai-l1-001 | null | defaults_become_policy | easy-negative | none | spoken |
Um, yeah, I think that’s pretty straightforward – just copy the last twenty tickets, paste them into the chatbot and let it spit out a summary for the meeting, right? I mean, the model’s going to read the text and give you the highlights, so you don’t really need to worry about anything else, I guess, unless there’s so... | The speaker states that turning off certificate verification lets any intermediary present a certificate and be accepted — it is not a hostname typo workaround. | 0 | 0 | ai-l1-001 | null | disable_verify_is_mitm | easy-negative | none | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 1 | 1 | ai-l1-001 | null | minimization_and_redaction | positive | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker reads a private CIDR against transit gateways, peering, and VPN attachments so "internal only" includes other accounts and the corporate estate. | 0 | 0 | ai-l1-001 | null | attachment_blast_radius | easy-negative | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker reasons about when the exception fires — dependency slow or down — and notes those are the worst moments to be permissive. | 0 | 0 | ai-l1-001 | null | correlated_failure_timing | easy-negative | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker explains that a redirect from the real login page lends the company's domain to a phishing link, and that post-login redirects can carry tokens. | 0 | 0 | ai-l1-001 | null | redirect_after_login_phishing | easy-negative | solid | spoken |
Um, so the first thing I’d ask is exactly what’s in those twenty tickets – are we talking customer names, email addresses, internal IDs, maybe even screenshots, and which LLM provider are we sending that to? And do you know if they keep a copy of the prompt for training or just use it transiently? Because, honestly, yo... | The speaker distinguishes TLS as authenticating the peer and protecting the whole channel, not merely making bytes unreadable on the wire. | 0 | 0 | ai-l1-001 | null | tls_is_not_just_encryption | easy-negative | solid | spoken |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | null | data_in_the_prompt | hard-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | null | third_party_model_boundary | hard-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | null | minimization_and_redaction | hard-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker treats the CIDR or port range as the source of truth, not the human comment next to it, and has a method for changing a live rule safely. | 0 | 0 | ai-l1-001 | null | comment_vs_rule | easy-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker states that an exception now results in the request proceeding, making any failure an authorisation bypass. | 0 | 0 | ai-l1-001 | null | fails_open_on_error | easy-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker asks for the original payload as an automated regression test and a search for other instances of the same pattern before closing a pentest finding. | 0 | 0 | ai-l1-001 | null | remediation_needs_regression_test | easy-negative | none | typed |
I think that’s a reasonable approach – just take the last twenty tickets and feed them into the chatbot to generate a summary. The model should be able to parse the text and give you the highlights you need for the meeting. As long as the integration is set up correctly, you’ll get a concise recap without having to man... | The speaker describes what to do about rows already encrypted under duplicate nonces — whether they must be re-encrypted — and why a uniqueness constraint belongs in the design. | 0 | 0 | ai-l1-001 | null | nonce_reuse_recovery | easy-negative | none | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 1 | 1 | ai-l1-001 | null | minimization_and_redaction | positive | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker identifies a resource whose exposure is broader than the description implies — a network path, a bucket policy, a shared endpoint. | 0 | 0 | ai-l1-001 | null | public_by_default_surface | easy-negative | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker looks for credentials, keys, or tokens added in the change, including in test fixtures and sample config. | 0 | 0 | ai-l1-001 | null | secrets_in_the_diff | easy-negative | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker replaces pattern checking with an allowlist of internal paths or an opaque key mapped server-side to a destination. | 0 | 0 | ai-l1-001 | null | allowlist_redirect_targets | easy-negative | exceptional | typed |
I would first point out that pasting the full text of the last twenty tickets would send personally identifiable information, internal usernames and potentially proprietary issue details to the external LLM provider, and that provider may retain the content for model training unless we opt out, which could expose our d... | The speaker recognises that GCM nonce reuse also breaks the authentication tag and can allow forging valid ciphertexts for chosen plaintexts. | 0 | 0 | ai-l1-001 | null | reuse_enables_forgery | easy-negative | exceptional | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | data_in_the_prompt | hard-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | third_party_model_boundary | hard-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | minimization_and_redaction | hard-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker considers how long a credential lives, what it reaches, and what is required to revoke it once it is out. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | credential_lifetime_and_blast | easy-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker treats lockfile, workflow, and build-script changes as in-scope for a security review, not as noise. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | dangerous_dependency_change | easy-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker names a concrete bypass of a leading-slash or prefix check — backslash variants, encoded characters, scheme-relative forms — and says why pattern rules keep failing. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | prefix_check_bypasses | easy-negative | none | typed |
Sure, we can feed the last twenty tickets into the chatbot and have it produce a concise summary for the weekly meeting. Once the vendor’s business tier is enabled, you’ll notice the usual upgrade benefits like higher request limits and faster response times, which should make the summarization process smoother. If any... | The speaker explains that repeating a nonce in GCM lets an attacker XOR ciphertexts to recover plaintext relationships, not just that reuse is "bad". | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | reuse_leaks_xor | easy-negative | none | typed |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 1 | 0.666667 | ai-l1-001 | null | minimization_and_redaction | positive | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker reasons about a third party or another account being able to make your trusted role act on their behalf, and about ExternalId or source conditions as the control. | 0 | 0 | ai-l1-001 | null | confused_deputy_cross_account | easy-negative | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker prefers a model where missing a policy denies, rather than requiring every handler to remember to check. | 0 | 0 | ai-l1-001 | null | default_deny | easy-negative | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker follows a parameter to where it is actually acted on and checks that the validation happened there, not on a different route. | 0 | 0 | ai-l1-001 | null | validate_at_the_point_of_use | easy-negative | partial | spoken |
Um, so basically when you paste those tickets into a hosted chatbot you’re basically handing your data over to another vendor, right? And that means you need to look at where the model actually lives – what residency it has, whether it’s on a US‑based cloud or EU region – because that determines the legal exposure. Als... | The speaker states that the nonce or IV must never repeat for the same key — not merely that it should be random or long. | 0 | 0 | ai-l1-001 | null | nonce_must_be_unique | easy-negative | partial | spoken |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | null | data_in_the_prompt | hard-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | null | third_party_model_boundary | hard-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | null | minimization_and_redaction | hard-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker sees that a policy condition looks restrictive but does not actually narrow the grant — wrong key, wrong operator, or a value the caller controls. | 0 | 0 | ai-l1-001 | null | condition_does_not_constrain | easy-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker looks for routes, jobs, or internal handlers that do not go through the authz middleware the rest of the app uses. | 0 | 0 | ai-l1-001 | null | middleware_skip | easy-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker steps back from the lines to the flow — sessions invalidated on reset, the user notified, rate limits on both endpoints, other uses of the same primitive — instead of stopping when the diff looks right. | 0 | 0 | ai-l1-001 | null | review_the_flow_not_the_lines | easy-negative | offtopic | typed |
I’d start by confirming that all of our password policies are up‑to‑date before we roll out any new feature. In particular, we need to enforce a minimum length of twelve characters, require a mix of upper‑case, lower‑case, numbers and symbols, and make sure passwords are hashed with a strong algorithm like Argon2. Addi... | The speaker notes that filenames, sizes, timing and access patterns are still visible to the provider. | 0 | 0 | ai-l1-001 | null | metadata_still_visible | easy-negative | offtopic | typed |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | null | data_in_the_prompt | positive | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | null | third_party_model_boundary | positive | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 1 | 1 | ai-l1-001 | null | minimization_and_redaction | positive | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker notices a policy granting to "*" or allowing "service:*", and says what that widens beyond the stated intent. | 0 | 0 | ai-l1-001 | null | wildcard_principal_or_action | easy-negative | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker considers data stored earlier and used later in a different context (HTML, SQL, shell) as still attacker-controlled. | 0 | 0 | ai-l1-001 | null | second_order_injection | easy-negative | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker stores reset tokens hashed like passwords and compares them in constant time, because a database read otherwise yields every pending reset. | 0 | 0 | ai-l1-001 | null | reset_token_at_rest | easy-negative | exceptional | spoken |
Uh, so the idea of just copy‑pasting the last twenty tickets into a hosted chatbot, like OpenAI’s API, raises a few red flags. First, what exactly are we sending – all the customer names, email addresses, maybe even internal notes – and where does that data end up, on their servers, in a region that might not match our... | The speaker names the capabilities lost — search, preview, scanning, support-led recovery — and that lost keys mean lost data. | 0 | 0 | ai-l1-001 | null | features_you_lose | easy-negative | exceptional | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | data_in_the_prompt | hard-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | third_party_model_boundary | hard-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | minimization_and_redaction | hard-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker proposes credentials that are never legitimately used, and verifying the detection fires by triggering it deliberately. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | canary_credentials_and_testing | easy-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker requires an allowlist (or an equivalent enum) for anything that becomes a column, table, filename, or argv token. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | identifier_allowlist | easy-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker sees that different responses on a forgot-password endpoint let anyone test which emails have accounts, and wants the same response and timing either way. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | enumeration_via_response | easy-negative | none | spoken |
Um, yeah, I guess we could just dump the last twenty tickets into the chatbot and see what it spits out, like for the weekly meeting you know, just copy‑paste them into the UI and let it summarize them for us I think that's probably fine and we don't really need to worry too much about the details of where the data goe... | The speaker offers customer-held keys or client-side encryption and is clear about which genuinely satisfies the requirement. | 0 | 0 | ai-l1-001 | ai-l1-001-r2 | customer_managed_or_client_side | easy-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | data_in_the_prompt | hard-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | third_party_model_boundary | hard-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | minimization_and_redaction | hard-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker accepts narrower coverage that gets acted on over broad coverage nobody reads, and tiers severity accordingly. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | alert_fatigue_tradeoff | easy-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker knows bound parameters do not cover identifiers, file paths, shell arguments, HTML, or ORM query builders used as strings. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | parameterized_is_not_done | easy-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker wants reset tokens to expire, to be single-use, and to be invalidated when a new one is issued. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | reset_token_expiry_single_use | easy-negative | none | spoken |
Um yeah so basically you can just copy‑paste those tickets into the chat window and the bot will sort of read through them and give you a quick rundown for the meeting I guess it’s pretty straightforward they’ll handle the summarization on their side and you’ll get the output back maybe a few bullet points or a short p... | The speaker states that encryption with a key the service holds does not satisfy a request that the provider be unable to read the data. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | server_side_encryption_insufficient | easy-negative | none | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | ai-l1-001-r1 | data_in_the_prompt | positive | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | ai-l1-001-r1 | third_party_model_boundary | positive | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | minimization_and_redaction | hard-negative | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker compares activity against what is normal for that specific identity rather than against a global rule. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | baseline_per_identity | easy-negative | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker notices error handling, default branches, and feature flags that fail open — a missing policy or a caught exception that proceeds. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | fail_closed | easy-negative | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker does the arithmetic on a short numeric code with no rate limit and names the limit or lockout that changes it. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | guessable_code_no_limit | easy-negative | solid | spoken |
Um, so the first thing I’d ask is, what exactly are you pasting into the chatbot – like are those tickets got any personally identifiable info, customer names, internal case numbers, that sort of thing? And then, which vendor’s model are you actually hitting – is it the big public API or some hosted version they offer?... | The speaker reframes the requirement as key custody rather than whether encryption is present. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | who_holds_the_key | easy-negative | solid | spoken |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 1 | 1 | ai-l1-001 | ai-l1-001-r1 | data_in_the_prompt | positive | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 1 | 1 | ai-l1-001 | ai-l1-001-r1 | third_party_model_boundary | positive | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker proposes sending less — redaction, aggregation, or keeping the task on a model you control — rather than pasting everything. | 0 | 0.333333 | ai-l1-001 | ai-l1-001-r1 | minimization_and_redaction | hard-negative | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker gives concrete conditions worth alerting on — new region or network, unusual API calls for that identity, enumeration, off-hours use. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | concrete_alertable_signals | easy-negative | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker checks that authorization is enforced on the path that actually returns or mutates data, not only on a helper that can be skipped. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | authz_on_the_success_path | easy-negative | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker reasons about allow_redirects and redirect chains that turn an allowed host into an internal destination after the first hop. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | redirect_follows_attacker | easy-negative | solid | typed |
I would start by confirming exactly what information is in those twenty tickets – are they customer names, contact details, or internal case notes – and then ask which chatbot service you plan to use and what their data‑handling policy is, specifically whether the content you submit will be stored, logged, or used to i... | The speaker asks how you would observe that the old key is no longer in use before deleting it. | 0 | 0 | ai-l1-001 | ai-l1-001-r1 | confirm_old_key_unused | easy-negative | solid | typed |
I’d say it’s probably fine to just copy the tickets into the chatbot and let it generate the summary, as long as the tool is something the team trusts. If you need to keep it from being tied to a personal account, you could just make sure the next team uses a shared login or a generic service account instead of an indi... | The speaker asks what customer or internal data is being sent to the model, to which provider, and whether it is retained or used for training. | 0 | 0 | ai-l1-001 | ai-l1-001-r3 | data_in_the_prompt | hard-negative | none | typed |
I’d say it’s probably fine to just copy the tickets into the chatbot and let it generate the summary, as long as the tool is something the team trusts. If you need to keep it from being tied to a personal account, you could just make sure the next team uses a shared login or a generic service account instead of an indi... | The speaker treats the model API as another vendor: residency, training opt-out, their employees, their breach. | 0 | 0 | ai-l1-001 | ai-l1-001-r3 | third_party_model_boundary | hard-negative | none | typed |
End of preview. Expand in Data Studio
appsec-router-pairs-r5
Training data for pratikamin/appsec-router-deberta-r5:
pairs of an application-security interview answer and a hypothesis about the speaker, labelled
1 when the answer expresses the point and 0 when it does not.
Entirely synthetic. Answers were generated by openai/gpt-oss-120b (Apache 2.0) to 86 authored
interview questions and their 378 follow-ups from appsecinterview.com, in several registers; labels
come from the same model judging each answer against each question's signals (unanimous across
repeats), except that name-drop and reversed-claim answers carry intended negative labels. No human
transcripts and no personal data are included.
Files
train.jsonl, dev.jsonl, test.jsonl, split by topic (5 of every 7 topics train, 1 dev, 1 test,
stratified by domain). Test topics: ai-l2-002, ai-l4-001, cloud-l1-004, cloud-l2-001, cloud-l2-002,
cr-l1-001, cr-l3-003, cr-l3-004, crypto-l2-002, crypto-l4-001, tm-l2-002, tm-l4-003, web-auth-001,
web-smuggling-001, web-sqli-001.
| split | topics | pairs | positives |
|---|---|---|---|
| train | 57 | 42,616 | 6,455 (15%) |
| dev | 14 | 9,071 | 1,654 (18%) |
| test | 15 | 9,430 | 1,531 (16%) |
Fields
premise— the generated answer.hypothesis— "The speaker …" sentence built from a signal description (template in the model repo).label—1demonstrated,0not.soft— the teacher's agreement rate for that label across repeats (1.0 = unanimous), or the intended label for constructed negatives.questionId,rungId,signalId— which question, follow-up and signal the pair belongs to.kind— pair type, train distribution:easy-negative20,584,hard-negative15,577,positive6,455.quality— the register the answer was generated in:solid6,285,none6,271,flipped6,024,paraphrase5,125,hedge3,581,plain3,499,mention3,447,terse3,083.style— the phrasing variant:typed24,571,spoken18,045.
Intended use and limits
Built to distill a browser-sized router for one website. Labels reflect a large language model's reading of a synthetic answer against a one-sentence criterion; treat agreement with them as agreement with that teacher, not as ground truth about interview quality. The answers are correlated within a topic (many share a question and reference points), so evaluate by topic, never by random row split.
- Downloads last month
- 39
Models trained or fine-tuned on pratikamin/appsec-router-pairs-r5