AI Safe Harbor & Report Quality (Avoiding the Spam Collapse)¶
As AI tooling accelerates vulnerability discovery, disclosure channels are increasingly stressed by high-volume, low-signal submissions.
This page sets a standard: responsible AI testing + high-quality reporting.
Safe Harbor (what it means)¶
A “good faith” safe harbor policy typically clarifies: - what testing is authorized - how data should be handled - how to disclose responsibly - how the org will respond legally
For AI systems, safe harbor often matters because: - model behavior is unpredictable - testing can touch sensitive data - legal ambiguity chills disclosure
Report quality bar (minimum)¶
If you file a report—AI-assisted or not—include:
- Clear scope + target
- Precise steps to reproduce (deterministic when possible)
- Evidence:
- request/response captures
- screenshots where relevant
- exact payloads
- Clear impact statement
- Suggested mitigations (if you can)
If you can’t reproduce it twice, it’s probably not ready.
AI-specific pitfalls¶
- Hallucinated endpoints / non-existent parameters
- “Vuln-shaped” output without proof
- Overstated impact
- Unsafe testing that violates policy or harms users
Defender guidance¶
- Provide explicit safe harbor language for AI research.
- Rate limit + require evidence.
- Use structured intake forms and require minimal PoC.
- Reward quality, not volume.
Source / inspiration¶
- Inspired by recent safe-harbor announcements for AI research and real cases where bug bounty programs were overwhelmed by low-signal submissions.