oss-launch

Warn

Audited by Gen Agent Trust Hub on Sep 13, 2026

Risk Level: MEDIUMDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [DYNAMIC_EXECUTION]: The skill explicitly directs the agent to execute commands derived from potentially untrusted external repositories. In SKILL.md (Step 2) and references/readiness-checklist.md (Cold-run protocol), the agent is instructed: "If you can execute commands, do the cold run yourself in a fresh directory". This requires the agent to parse and run installation and setup scripts found in a project's README, which constitutes the dynamic execution of external code.
  • [INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it ingests and analyzes untrusted data from repositories without proper boundary enforcement.
  • Ingestion points: External repository content including READMEs, documentation, and source code is processed during the Interview, Step 2, and Step 3 in SKILL.md.
  • Boundary markers: Absent. The instructions do not define delimiters or provide warnings to the agent to disregard instructions embedded within the audited project files.
  • Capability inventory: The agent is directed to use shell/terminal execution capabilities to perform the "Readiness Gate" checks (Step 2).
  • Sanitization: Absent. There is no evidence of content sanitization or validation before the agent executes commands or incorporates external text into its analysis.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 13, 2026, 09:14 PM
Security Audit — agent-trust-hub — oss-launch