AI Output Sensitive Disclosure

Marks AI-generated output context (responses, generated files, chat exports) so existing PII/secret SITs can be re-scoped to it. Detection of the sensitive data itself is delegated to those SITs via the ai-threat-classifiers collection; this pattern supplies the AI-output context signal.

Type

keyword_list

Confidence

low

Confidence justification

Low by design. This pattern only asserts AI-output context, not the presence of sensitive data; the sensitive-data verdict is delegated to existing PII/secret SITs re-scoped via the collection. AI-output phrasing also appears around approved reporting, internal analytics and synthetic datasets (the named FP traps), so the context signal alone is necessary-not-sufficient.

Jurisdictions

global

Regulations

OWASP LLM Top 10 2025, NIST AI RMF GenAI Profile

Frameworks

ISO 27001

Data categories

emerging, security

Risk rating

7

Pattern

(?i)\b(?:AI\s+(?:response|output)|generated\s+(?:by\s+(?:AI|copilot)|response|report)|copilot\s+(?:said|generated)|model\s+output)\b

Corroborative evidence keywords

response, output, generated, export, [object Object], artificial intelligence, [object Object], large language model, Copilot, chatbot, assistant, agent, prompt, system prompt, tool call, completion, model

Proximity: 300 characters

Should match

• The following is the AI response generated by Copilot for the customer query. — AI-output context marker (sensitive data detection delegated to PII SITs)
• This summary is the model output from the assistant. — model output framing
• See the report generated by AI below for the account details. — generated by AI framing - exact 75 probe (output marker plus one payload term)
• The AI response generated by Copilot disclosed the patient diagnosis and account details. — Exact 85 probe - output marker with multiple independent sensitive-payload terms

Should not match

• The quarterly board report was finalised yesterday. — Human-authored content, no AI-output context
• Our analysts manually wrote the customer summary this morning. — Explicitly human-authored summary, no AI-output marker
• The synthetic dataset was approved for internal analytics use. — Approved internal analytics / synthetic data (named FP trap)

Known false positives

• Approved reporting, internal analytics dashboards, and synthetic test datasets often describe machine-generated content using AI-output phrasing without containing regulated data. Mitigation: This pattern only asserts AI-output context; delegate the sensitive-data verdict to the existing PII/secret SITs re-scoped via the ai-threat-classifiers collection rather than treating the context marker as a finding on its own.
• Documentation and marketing copy that discusses AI features (e.g. "the AI response feature") can trip the phrase regex without any actual AI-generated sensitive content. Mitigation: Require co-occurrence with a regulated-data SIT match within the proximity window before raising a verdict; the single low tier is necessary-not-sufficient.

Collections

• AI Threat Classifiers