software-data-integrity-failures

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: Auditing external code introduces an injection surface where untrusted data is processed. • Ingestion points: User-provided Python source code and templates in SKILL.md and references/deserialization-exploits.md. • Boundary markers: Extensive use of Markdown headers, checklist items, and explicit Safe/Unsafe labels to distinguish between documentation and vulnerable code patterns. • Capability inventory: The skill does not define external tool access, providing instead Python AST patterns and shell grep commands for the agent to use locally. • Sanitization: Relies on underlying platform guardrails for processing external content.
  • [COMMAND_EXECUTION]: The skill identifies and documents dangerous command execution patterns (e.g., os.system, subprocess.run) to teach users how to audit and secure their own applications.
  • [REMOTE_CODE_EXECUTION]: The skill provides documentation on remote code execution vulnerabilities via pickle deserialization and unverified network downloads as part of its security auditing educational content.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 08:14 PM
Security Audit — agent-trust-hub — software-data-integrity-failures