deliver-retro

Pass

Audited by Gen Agent Trust Hub on Nov 12, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze transcripts and reports from prior agent sessions, which are considered untrusted data sources.
  • Ingestion points: Accesses orchestrator transcripts in ~/.claude/projects/ and ~/.codex/sessions/, as well as worker reports and plan versions in the run's reports/ directory.
  • Boundary markers: The instructions do not define specific delimiters or guardrails to prevent the agent from following malicious instructions that might be embedded in the logs it is reviewing.
  • Capability inventory: The skill generates a local ledger (JSON) and offers to write TODO entries to an external project tracker via a shared procedure.
  • Sanitization: The skill mitigates risks by explicitly instructing the agent to 'retain sanitized pointers rather than raw transcript excerpts or secrets' and to 'print environment variable names or hashes, never values'.
  • [DATA_EXFILTRATION]: The skill facilitates the transfer of analyzed data from local logs to external project management tools.
  • Evidence: Step 9 of SKILL.md describes an 'outstanding-work procedure' to search for and create/update entries in a 'project's chosen tracker'.
  • Controls: This operation is conditional on explicit user confirmation and selection of specific items.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 12, 2026, 05:36 AM