The 90-Day Disclosure Standard
What you'll learn: How the 90-day disclosure timeline works for AI vulnerabilities, including grace periods and negotiating extensions for complex issues.
The Standard Timeline
When you discover a vulnerability in an AI system, the 90-day disclosure standard gives organizations roughly three months to fix it before you publicly disclose the issue. This timeline, borrowed from traditional cybersecurity, balances two competing needs: giving developers time to patch vulnerabilities while preventing indefinite suppression of critical safety information.
The clock typically starts when you submit a verified vulnerability report through the appropriate disclosure channel (as covered in lesson 3524). For straightforward issues—like a prompt injection that extracts training data—90 days is usually sufficient. However, AI systems often present unique complexity.
Grace Periods and Extensions
Unlike a simple software bug, AI vulnerabilities may require:
- Model retraining (computationally expensive, weeks of GPU time)
- Dataset remediation (identifying and removing problematic data)
- Alignment testing (ensuring fixes don't break other behaviors)
- Deployment coordination (updating models across distributed systems)
Responsible disclosure frameworks allow negotiated extensions when vendors demonstrate good-faith progress. You might grant an additional 30-60 days if the organization shows evidence of active remediation—documented retraining runs, partial patches, or coordinated deployment plans.
Think of it like renovating a building: you can't just slap a Band-Aid on structural issues. Complex AI fixes require architectural changes.
When the Deadline Matters
If 90 days (plus any negotiated extension) passes without resolution, you proceed to public disclosure—even if the issue remains unfixed. This deadline prevents indefinite delays and ensures the broader community learns about persistent risks. However, you should always weigh potential harm: actively exploited vulnerabilities affecting vulnerable populations might warrant earlier or coordinated disclosure.
Key Takeaway: The 90-day standard provides a reasonable timeline for most AI vulnerability fixes, but complex issues may require negotiated extensions with documented progress—never indefinite silence.