prompt-safety-review

Trace untrusted text into prompts and verify sanitization, prompt-injection regression coverage, and preservation of system-level instructions.

SKILL.md

GitHub

Prompt Safety Review

Workflow

  1. List all untrusted sources: market data strings, user input, model outputs, prior-agent reasoning, DB text, logs, and external API responses.
  2. Trace each source into prompt construction or agent context injection.
  3. Verify _sanitize_external_text() or an equivalent sanitizer runs before LLM injection.
  4. Confirm new prompt templates do not weaken system, safety, or risk-gate instructions.
  5. Check _PROMPT_INJECTION_PATTERNS coverage when new attack strings are introduced.
  6. Require regression tests for prompt-injection paths that changed.

Blockers

  • Raw cross-agent reasoning injected into a prompt.
  • External text appended to system instructions without sanitization.
  • Security tests that only assert the happy path.