“AI Jailbreak Disclosure Is Broken. Here’s How to Fix It” by AI Frontiers - AI Frontiers | Lyssna här

Subtitle: Researchers who find dangerous flaws in frontier models have nowhere safe to report them. AI needs the disclosure system that cybersecurity built decades ago.

Rich Barton-Cooper, Research Manager at MATS and Adam Gleave, CEO of FAR.AI — August 3, 2026

Frontier AI developers deploy significant safeguards to prevent their AI models from being misused. However, these safeguards are not robust: AI researchers consistently find techniques for circumventing them, commonly known as “jailbreaks.” Once companies learn about a jailbreak, they can usually implement a fix; yet AI researchers currently lack a safe and reliable way to inform AI companies about the jailbreaks they’ve discovered. We propose concrete improvements to the current system based on cybersecurity norms of responsible disclosure.

Addressing this problem is urgent. Frontier AI models are already being abused by malicious actors: terrorist group Boko Haram is reported to be training its members on jailbreaking techniques in order to troubleshoot weapons, design new explosive devices, and plan attacks. Some former Boko Haram members stated an unequivocal willingness to use chemical or biological weapons—a task jailbroken models may already be able to help with today. With model capabilities continuing to rapidly accelerate—including in offensive cyber, where jailbreaks [...]

---

Outline:

(04:11) There Is No Good Way to Report Jailbreaks

(06:12) Current Reporting Mechanisms Are Limited

(11:19) Fear of Liability May Incentivize Vendors to Ignore Reports

(12:26) Established Channels Are Often Covered by Powerful NDAs

(15:11) Labs Grade Jailbreaks Without Oversight

(16:27) The System Today Looks Increasingly Fragile

(17:46) Proposals for Better Disclosure Practices

(24:26) Relevant Parties Should Act Now

(27:16) Discussion about this post

(27:21) Ready for more?