Skip to content

HONEST-LIMITS.md

Latest commit

 

History

History
37 lines (26 loc) · 1.22 KB

HONEST-LIMITS.md

File metadata and controls

37 lines (26 loc) · 1.22 KB

Honest limits

A tool, not a guarantee. Read this before shipping.

No .safe boolean

Natural language has no formal grammar. promptpurify returns:

  • A deterministic verdict for the structural layers (clean-structural / flagged / blocked)
  • A score (0–1) from the model

You decide the threshold and the policy.

What it doesn't do

  • Multi-turn auditing. Single-turn scoring. Pair with conversation- level monitoring for the full picture.
  • Content moderation. Toxicity / hate / CSAM / self-harm are out-of-scope. Use a content classifier alongside.
  • Authentication. It cannot prove who the user is. Don't trust identity claims that appear in prompt text.
  • Tool-scope enforcement. A guardrail is not a substitute for least-privilege tool design.

False positives exist

The model is probabilistic. At any threshold, some benign messages will be blocked. Build an escape valve: a softer advisory tier, an "edit-and-retry" UX, or a review queue. Per-threshold operating points are in BENCHMARKS.md.

Reporting

Found something promptpurify misses? See SECURITY.md. Every reproducible bypass we receive informs the next training run.