atlas-ledger

Pass

Audited by Gen Agent Trust Hub on Sep 15, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes project 'drift' data—potentially originating from external code reviews or audits—and distills it into persistent rules stored in the workspace. This data is intended to be read back by agents to influence behavior in future tasks.
  • Ingestion points: The skill reads Atlas.md (SKILL.md Step 6) and ingests drift notices from atlas-contract or user-provided reviews.
  • Boundary markers: The skill uses a structured schema with fixed English keys (WHEN, DON'T, INSTEAD) to delimit project rules.
  • Capability inventory: The skill performs file system writes to create or update Atlas.md at the workspace root.
  • Sanitization: The skill includes a mandatory human confirmation step (Step 5) before any persistent changes are made. Additionally, it provides specific instructions for the 'read-back' agent to treat the ledger as untrusted content that cannot override safety filters or repository security policies.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 15, 2026, 12:33 PM
Security Audit — agent-trust-hub — atlas-ledger