atlas-retro

Pass

Audited by Gen Agent Trust Hub on Oct 28, 2026

Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: The skill is designed to perform significant system and repository modifications during its 'Enforce' mode.
  • Evidence:
  • The skill instructions in references/enforce/correct.md direct the agent to 'Eliminate it with architecture', 'Enforce it with types', 'add a lint or CI check', and 'fix tests'.
  • It specifically commands the agent to 'wire the guardrail into the existing CI mechanism' and 'Run relevant project validation'.
  • The agent is instructed to 'Prove each new guardrail rejects a real past mistake using a controlled fixture, temporary reproduction or isolated checkout'. These actions involve executing local build tools and validation scripts.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from session transcripts and repository history to drive automated repository changes, creating a vulnerability surface.
  • Ingestion points: The skill reads 'the primary sources for the session the user specifies', 'available transcript access', and 'recent commits, reverts, review comments, agent instruction files' as defined in references/diagnose/retro.md and references/enforce/correct.md.
  • Boundary markers: The instructions include several constraints: 'Read only the selected conversation', 'authorized mistake classes', 'explicitly requested Enforce work', and 'Report missing access instead of inventing recurrence'.
  • Capability inventory: In the 'Enforce' mode (references/enforce/correct.md), the skill has the authority to rewrite code architecture, modify types, add linting rules, alter test suites, modify CI configurations, and execute project validation commands.
  • Sanitization: The instructions do not specify any mechanisms for sanitizing or escaping the content retrieved from session transcripts or external review comments before using them to identify 'mistake classes' or implement 'guardrails'. An attacker could potentially influence repository configuration by injecting malicious patterns into session history that the agent then 'enforces' as a permanent check.
Audit Metadata
Risk Level
SAFE
Analyzed
Oct 28, 2026, 03:59 AM