atlas-mode
Pass
Audited by Gen Agent Trust Hub on Sep 19, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructs the agent to execute
npm run checkafter performing file modifications. While this is a common development practice for verifying changes, it involves direct shell command execution. - [INDIRECT_PROMPT_INJECTION]: As an orchestration skill, the agent ingests external "facts" and incorporates them into instructions for subagents. This processing of potentially untrusted data creates a vulnerability surface where malicious inputs could influence downstream tasks.
- Ingestion points: Step 2 in
SKILL.md("Gather planning facts") is the primary point where external information enters the agent's context. - Boundary markers: The skill requires the agent to use a specific "Handover format" and to "Mark guesses as assumptions" to differentiate between source data and reasoning.
- Capability inventory: The skill utilizes the
Agentandworkflowtools to spawn subagents and can execute shell commands (npm run check). - Sanitization: The skill mandates running a
/unslopprocess in Step 6 to clean prose and ensure generated tasks are concise and verifiable.
Audit Metadata