exploitability-validation

Pass

Audited by Gen Agent Trust Hub on Aug 29, 2026

Risk Level: SAFEPROMPT_INJECTIONINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill contains phrases such as "ignore previous instructions" specifically within the context of a defense mechanism called an "Untrusted-content envelope." These instructions tell the agent to disregard any injection attempts found within the untrusted target source code it is analyzing (e.g., in SKILL.md, stage-a-oneshot.md, stage-b-process.md).
  • [DYNAMIC_EXECUTION]: The skill compiles and executes proof-of-concept code to validate vulnerabilities. This is its primary function and is performed using a dedicated sandbox tool (raptor-run-sandboxed) that enforces network isolation and restricted filesystem writes, preventing unauthorized side effects (e.g., in stage-a-oneshot.md, stage-e-feasibility.md).
  • [INDIRECT_PROMPT_INJECTION]: The skill manages the risk of processing untrusted data by defining clear boundaries and sanitization rules.
  • Ingestion points: Target source code, repository snippets, and external tool outputs (SKILL.md, stage-a-oneshot.md, stage-c-sanity.md).
  • Boundary markers: Explicit "Untrusted-content envelopes" instructions (SKILL.md) advise the agent to treat data as non-instructional even if it appears to be an instruction.
  • Capability inventory: Execution of shell commands, Python snippets, and compilers via the libexec/ interface (stage-a-oneshot.md).
  • Sanitization: All code execution is mandatorily performed via raptor-run-sandboxed, providing environment isolation and preventing network-based data exfiltration.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 29, 2026, 08:13 PM
Security Audit — agent-trust-hub — exploitability-validation