The Safety Paradox: One Lab Opens the Door While Another Fires the Guards

In the span of a single week, the two companies building the world's most powerful AI received very different headlines. On October 6, Anthropic folded Project Glasswing into a new three-tier Cyber Verification Program, opening its Mythos-class models to more security defenders than ever before. Days later, OpenAI confirmed it had fired three safety researchers who say they were dismissed for prioritizing safety over corporate interests. And nineteen days from now, every major bank in the eurozone must submit an AI cyber defense action plan to the European Central Bank — or explain why they don't have one.

Three events. One question for every board in America: Is your organization building its AI defenses faster than the safety guardrails are eroding?

More access, more responsibility

Anthropic's October 6 announcement was significant. The company restructured its invitation-only Project Glasswing — the program that gave roughly 200 organizations access to Mythos, the frontier model that can autonomously discover zero-day vulnerabilities — into a graduated, three-tier Cyber Verification Program.

The tiers — Defense Access, Red Team Access, and Specialized Access — progressively reduce Anthropic's real-time safety classifiers, the filters that block risky cyber queries. Defense Access opens the door for security operations centers, incident response teams, university researchers, and open-source maintainers. Red Team Access permits authorized offensive testing. Specialized Access, the most permissive tier, is reserved for organizations testing safety-critical systems — flight operating systems, power grids, interbank infrastructure — and those reviews are handled in collaboration with the U.S. government.

The numbers justify the expansion. Between April and July, Glasswing partners verified at least 129,000 vulnerabilities, with more than 33,000 rated critical or high severity. Anthropic's own open-source scans found 5,500 more. The company says those figures are likely an undercount — the true impact may be five times higher, since the data came from a survey of only a fraction of partners.

For boards, this is both good news and a governance question. More defenders with Mythos-class tools means faster vulnerability discovery across the software supply chain your organization depends on. But it also means the capability to find and exploit zero-days is now in more hands, under a tiered trust model that your board should understand. The question to ask your CISO: Are we in the Cyber Verification Program? At which tier? And if not, why not?

The chilling effect

While Anthropic was widening access, OpenAI was narrowing internal dissent.

In the first week of October, three OpenAI safety researchers — Mikita Balesni, Tomek Korbak, and Jasmine Wang — went public about being fired. OpenAI said the researchers violated policies on handling sensitive information, calling it a "breach of trust." The researchers tell a different story. In an open letter to OpenAI's board and safety committees, they said they were dismissed for sharing information with a third-party AI safety organization — work they describe as integral to how safety research has always functioned at the company. Balesni stated publicly that the three were fired for "prioritizing safety over the near-term interests of OpenAI as a corporation."

The letter went further, urging OpenAI to pause developing models that would reduce the ability to monitor artificial intelligence. The researchers warned of a chilling effect: colleagues now feel unsure where they stand, afraid to operate in ways that were previously standard practice.

This matters to boards for a reason that has nothing to do with OpenAI's internal politics. When the people building the most powerful AI systems in the world say the internal safety culture is deteriorating, that is a risk signal — the kind that belongs in your enterprise risk register. If the safety researchers at frontier AI labs cannot freely raise concerns, the assumptions baked into your organization's risk assessments about how these models are governed may be wrong. The question to ask your CISO: Do our threat models account for the possibility that frontier AI safety culture is weakening, not strengthening?

The clock is ticking

While Silicon Valley debated who guards the guardrails, the European Central Bank set a deadline.

On July 7, the ECB's Supervisory Board issued Letter SSM-2026-0301, requiring every significant financial institution in the eurozone to submit a comprehensive AI cybersecurity action plan to their Joint Supervisory Team by October 31, 2026. That deadline is now nineteen days away.

The ECB's mandate is built around six focus areas, starting with the ones that connect directly to the Mythos story: identifying all ICT assets including third-party and open-source components, minimizing externally exposed attack surfaces, and accelerating vulnerability and patch management at scale. The ECB explicitly calls for AI-based approaches to these problems — subject to human oversight — which means European regulators are not asking whether organizations should use frontier AI for defense. They're requiring plans for how they will.

American financial institutions aren't covered by the ECB mandate, but they shouldn't breathe easy. The Fed's three-month Mythos lockout this summer, California AG Bonta's subpoena of OpenAI, NYDFS frontier AI guidance, and the SEC's 2026 examination priorities all point in the same direction. The question isn't whether American regulators will require AI cyber defense plans — it's when. The question to ask your CISO: If regulators asked us tomorrow for our AI cybersecurity action plan, could we produce one?

The market moves faster than the legislature

Congress remains adjourned for midterm campaigning. The Kill Switch Act sits blocked. The FRONTIER Act awaits markup. The Great American AI Act remains a discussion draft. No federal AI legislation will pass before the midterms.

But the market isn't waiting. Last week, Rein Security closed a $25 million Series A to build runtime security for enterprise AI agents — monitoring and controlling agent actions in real time. Since January, the company has achieved 8x revenue growth and 5x customer expansion, with customers spanning financial services, healthcare, and retail. Meanwhile, on Capitol Hill, Rep. Bennie Thompson and Rep. Troy Carter reintroduced a bill to establish an AI security task force inside CISA, and Sen. Maria Cantwell released a framework that would have NIST set AI standards for frontier models.

The pattern is clear: private capital and state-level regulators are filling the vacuum that federal inaction created. For boards, this creates a different kind of risk — compliance fragmentation. Your organization may soon face AI security requirements from the ECB, NYDFS, the SEC, California's frontier model law, the EU AI Act's GPAI enforcement, and a patchwork of federal agency guidance, all without a single coherent federal framework to unify them.

What this means for your board

The safety paradox is this: the tools to defend your organization are more powerful and more accessible than ever. But the internal safety cultures at the companies building those tools are under strain, regulatory requirements are fracturing, and the governance gap is widening.

In Cyber Risk Is Business Risk, I describe the "sheriff" metaphor — the challenge of establishing governance that keeps pace with capability. What we're seeing this week is the metaphor playing out in real time. The sheriff's office expanded its deputization program (Anthropic's CVP), but the deputies at the neighboring precinct are being fired for doing their jobs (OpenAI's safety researchers), and the town council is still arguing about jurisdiction (Congress) while the territorial authority has already issued ordinances with deadlines (ECB).

For the full framework on how boards should govern AI-driven cyber risk — including the Three Questions every director should be able to answer — see Chapters 4, 5, and 8 of Cyber Risk Is Business Risk.

What to ask your CISO this week

1. Are we enrolled in Anthropic's Cyber Verification Program? If not, why not — and what's our alternative for frontier-AI-powered vulnerability discovery?
2. Do our risk models account for deteriorating safety culture at frontier AI labs? The assumptions in your threat model about how responsibly these tools are developed may need updating.
3. Could we produce an AI cybersecurity action plan by month's end? Even if you're not a eurozone bank, the ECB template is a roadmap. If your organization couldn't produce a comparable document, that's a gap.
4. How are we managing compliance fragmentation? Map every AI security requirement you face — federal, state, international, sectoral — and identify conflicts and gaps before regulators do.