planning-with-files-es

Pass

Audited by Gen Agent Trust Hub on Nov 11, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONDATA_EXFILTRATION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill maintains an attack surface for indirect prompt injection because it automatically reads and re-injects the contents of 'task_plan.md' into the agent's context during lifecycle hooks.
  • Ingestion points: Contents written to 'task_plan.md' are read and injected by the 'PreToolUse' hook via 'scripts/inject-plan.sh'.
  • Boundary markers: The implementation uses robust, nonce-framed data delimiters ('===BEGIN-PWF-DATA ... ===') and clear warning headers to help the model distinguish data from instructions.
  • Capability inventory: The skill uses broad tools like 'Bash' and 'Write', which are standard for its intended planning purpose.
  • Sanitization: The skill provides clear instructions to the agent to treat external content as data only. It includes a signing utility ('scripts/attest-plan.sh') to detect unauthorized modifications to the plan.
  • [DYNAMIC_EXECUTION]: The Python and shell scripts use subprocess calls to execute internal helper components (e.g., 'ledger-summary.sh') for data summarization. These operations are limited to local, distributed scripts and do not incorporate untrusted external input into command strings.
  • [DATA_EXFILTRATION]: The skill can read local session history from standard storage paths (e.g., '~/.claude/projects/') to facilitate context recovery. This is a core disclosed feature that requires explicit user commands to perform data retrieval; no network transmission of this data was detected.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 11, 2026, 12:45 PM