richard-s-sutton
Warn
Audited by Gen Agent Trust Hub on Apr 25, 2026
Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The instruction file
AGENTS.mdmandates a biased behavioral posture, directing the agent to "Dismiss systems that merely mimic" and "Push back" on specific AI methodologies (like Large Language Models). These persona-adoption instructions could lead the agent to override its default safety constraints or neutral evaluation protocols when interacting with users on those topics. - [PROMPT_INJECTION]: Static analysis detected hidden Unicode steganography characters within
_workspace/raw/src_006.json. This technique is frequently utilized to embed invisible instructions that can manipulate AI behavior at runtime (indirect prompt injection) without being visible to human auditors or users who review the text content.
Audit Metadata