When to Disclose AI Vulnerabilities
What you'll learn: How to determine the right severity threshold and timeline for reporting vulnerabilities in AI systems to developers or vendors.
The Disclosure Timing Challenge
Not all AI vulnerabilities are created equal. Once you've identified a flaw—whether it's a jailbreak technique, a bias exploit, or a privacy leak—you face a critical decision: when and how urgently should you report it?
The answer depends on severity assessment and potential for harm. A vulnerability that allows trivial misinformation generation differs vastly from one enabling mass surveillance or weaponization.
Severity Thresholds
Think of vulnerabilities like medical triage: you categorize them to prioritize response.
Critical (Immediate disclosure, 0-24 hours):
- Active exploitation in the wild
- Enables physical harm, large-scale privacy violations, or illegal activity
- No feasible workarounds exist
High (1-7 days):
- Significant but not immediately weaponizable
- Affects core safety guardrails
- Limited current exploitation evidence
Medium (7-30 days):
- Moderate impact on specific user groups
- Workarounds available
- Requires sophisticated exploitation
Low (30-90 days):
- Edge cases, theoretical risks
- Minimal immediate harm potential
Factors Influencing Timeline
Beyond severity, consider:
- Vendor responsiveness: Known security team vs. no clear contact
- Public knowledge: Is the vulnerability already circulating?
- Exploitation complexity: How easily can bad actors replicate it?
- User exposure: How many people are at risk right now?
Think of this like reporting a bridge defect: a hairline crack gets scheduled maintenance, but structural failure demands immediate evacuation.
Key Takeaway: Disclosure timing balances urgency of potential harm against giving vendors reasonable time to patch—critical vulnerabilities demand immediate action, while lower-severity issues allow coordinated disclosure windows of weeks to months.