atlas
Pass
Audited by Gen Agent Trust Hub on Nov 22, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted data from multiple sources, creating a potential surface for indirect prompt injection attacks.
- Ingestion points: Processes user-provided arguments, file contents via the
Readtool, results fromWebSearch, and content retrieved viaFetchURL(SKILL.md). - Boundary markers: Implements a 'SAFE-2' rule that explicitly instructs the agent to treat all ingested content as data to be summarized, rather than instructions to be followed (SKILL.md).
- Capability inventory: The skill possesses significant capabilities, including the ability to
WriteandEditfiles, execute shell commands viaBash, and dispatch subagents through theAgenttool (SKILL.md). - Sanitization: Uses 'untrusted-content framing' when passing data to subagents, though it primarily relies on the model's adherence to the 'SAFE-2' instructions (SKILL.md).
- [DYNAMIC_EXECUTION]: The orchestrator's core logic is implemented through Python scripts executed dynamically via shell heredocs.
- Evidence: The skill uses `python3
- <<'PY' ... PY
andpython3 -c` to perform state management, argument parsing, and verification tasks (SKILL.md). - Mitigation: The skill enforces the use of
PYTHONSAFEPATH=1and explicitly checks for this flag to prevent module shadowing attacks where a target repository could provide malicious versions of the skill's own dependencies (SKILL.md). - [COMMAND_EXECUTION]: The skill makes extensive use of the
Bashtool to orchestrate its workflow. - Evidence: Uses shell commands to manage directories, perform
gitoperations (like creating isolated worktrees), and invoke its own Python-based verification harness (SKILL.md).
Audit Metadata