audit-archived-plan-retrospectives

Pass

Audited by Gen Agent Trust Hub on Jul 11, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to the ingestion of untrusted external data (logs, quality findings, and user-generated plan artifacts) which are processed to generate lessons and reports.
  • Ingestion points: Files processed include logs/script-execution.log, logs/work.log, artifacts/findings/*.jsonl, and references.json.
  • Boundary markers: The skill explicitly instructs the agent to surface script-computed results 'verbatim' and forbids paraphrasing or re-deriving logic inline, which acts as a delimiter between untrusted data and agent reasoning.
  • Capability inventory: The agent has access to Bash, Write (to create lesson files and move plan directories), and AskUserQuestion.
  • Sanitization: The skill utilizes a deterministic Python script (scripts/audit.py) to handle data parsing and computation, reducing the LLM's direct exposure to raw adversarial prose during the analysis phase.
  • [COMMAND_EXECUTION]: The skill executes local Python scripts (scripts/audit.py and .plan/execute-script.py) to perform its core functions. These are standard within the plan-marshall platform architecture and do not involve remote code downloads or unverified external dependencies.
Audit Metadata
Risk Level
SAFE
Analyzed
Jul 11, 2026, 07:12 AM
Security Audit — agent-trust-hub — audit-archived-plan-retrospectives