AI safety researchers who find critical vulnerabilities in advanced AI models have nowhere safe to report them, creating a dangerous gap in how the industry handles security risks. Unlike cybersecurity, which developed formal disclosure protocols decades ago, the AI field lacks a standardized system for researchers to confidentially report dangerous flaws to developers and regulators. This absence leaves potentially catastrophic vulnerabilities unaddressed and researchers vulnerable to legal liability.

What Happens When Researchers Find Dangerous AI Flaws?

When security researchers discover what's called a "jailbreak" (a method to bypass an AI model's safety guardrails), they face a difficult choice. They can report the flaw directly to the AI company that built the model, but without a formal agreement protecting them, they risk legal action. They can publish their findings publicly, which alerts the world to the vulnerability but also gives bad actors a roadmap. Or they can stay silent, which protects them legally but leaves the flaw unpatched.

This problem is especially acute for frontier AI models, the most advanced systems built by companies like Anthropic, OpenAI, and Google DeepMind. These models are powerful enough that vulnerabilities could have serious real-world consequences, yet the industry has no agreed-upon way to handle disclosure.

How Should AI Companies Handle Security Reports?

The cybersecurity industry solved this problem in the 1990s and 2000s through coordinated disclosure frameworks. When a researcher finds a vulnerability in software, they report it to the vendor under a non-disclosure agreement (NDA). The vendor gets time to patch the flaw before the researcher can publish details publicly. This protects users while giving researchers legal cover and credit for the discovery.

AI companies need a similar system, but with modifications for the unique challenges of AI safety. The field requires several key components to make disclosure work:

• Universal Jailbreak Standards: A shared definition of what counts as a dangerous flaw, so researchers and companies agree on severity and urgency.
• Third-Party Clearinghouses: Independent organizations that can receive reports, verify them, and coordinate between researchers and developers without favoring any single company.
• Jailbreak Severity Rubric: A standardized scale for rating how dangerous a vulnerability is, similar to the CVSS (Common Vulnerability Scoring System) used in cybersecurity.
• Bug Bounty Programs: Financial incentives for researchers to report flaws responsibly rather than selling them to bad actors or publishing them publicly.
• Legal Protections: Clear agreements that protect researchers from liability when they report in good faith.

Without these structures, the current system creates perverse incentives. Researchers who want to do the right thing face legal risk. Companies that want to fix flaws can't easily find out about them. And the public remains exposed to vulnerabilities that could be patched.

Why Does This Matter for AI Risk?

The stakes are particularly high for frontier AI models because these systems are becoming more capable and more widely deployed. A jailbreak that allows an AI model to bypass its safety guidelines could enable harmful uses, from generating malware code to impersonating people at scale. The more advanced the model, the more dangerous an unpatched vulnerability becomes.

The absence of a disclosure system also undermines broader AI safety efforts. Researchers who study AI risks need to be able to report their findings without fear of legal retaliation. Companies need to know about flaws so they can improve their models. Regulators need visibility into what vulnerabilities exist so they can set appropriate standards. Right now, all of that information stays hidden.

The good news is that the solution already exists in another industry. Cybersecurity has spent three decades building disclosure norms, legal frameworks, and organizational structures that work. AI can adapt those lessons rather than reinventing the wheel. The question is whether the industry will act before a major vulnerability slips through the cracks.