Anthropic’s New Cybersecurity Model Could Get It Back in the Government’s Good Graces

On April 7, 2026, Anthropic pulled off something remarkable in the AI industry: it announced a breakthrough model and then deliberately refused to let anyone use it. The model, called Mythos, is so powerful at finding and exploiting software vulnerabilities that Anthropic deemed it “too dangerous for public release” — the first time a major AI developer has withheld a system since OpenAI temporarily held back GPT-2 in 2019.

But Anthropic didn’t just lock Mythos away. Instead, it launched Project Glasswing, a carefully curated initiative giving a select group of organizations — including Microsoft, Google, Apple, Amazon Web Services, JPMorgan Chase, and Nvidia — controlled access to the model for defensive cybersecurity testing. The strategy is as much about politics and public relations as it is about security: by positioning Mythos as a tool for protection rather than attack, Anthropic is attempting to rebuild trust with government regulators who have grown increasingly wary of Big Tech’s AI ambitions.

The Capabilities That Made Headlines

Anthropic accompanied the Mythos announcement with a 245-page technical document detailing what the company describes as a major leap in AI capability. The model operates at the level of a senior software engineer, demonstrating the ability to spot subtle bugs and self-correct mistakes in real time. On the USAMO 2026 Mathematical Olympiad — a grueling, two-day proof-based competition — Mythos scored 31 percentage points higher than Anthropic’s previous cutting-edge model, Opus 4.6.

That same coding prowess, however, makes Mythos a formidable offensive weapon. In internal testing, the model found critical vulnerabilities in every widely used operating system and web browser. Of those vulnerabilities, 99 percent have not yet been patched. Anthropic has disclosed only a fraction of what it claims to have discovered.

Independent evaluations confirm the threat is real, even if more bounded than Anthropic’s press materials suggest. The U.K.’s AI Security Institute (AISI), which was granted early access, found that Mythos succeeded in expert-level hacking tasks 73 percent of the time. Before April 2025, no AI model could complete those tasks at all.

“The fallout — for economies, public safety, and national security — could be severe.” — Anthropic’s official statement

Project Glasswing: A Controlled Rollout

Rather than a public release, Anthropic is limiting Mythos access through Project Glasswing to a small group of organizations that can use it defensively. The concept is straightforward: let trusted partners scan their own networks, discover vulnerabilities before bad actors do, and patch problems quietly.

The initial cohort reads like a who’s-who of the technology and financial sectors:

Microsoft, Google, Apple — the three companies whose operating systems and browsers were all found vulnerable by Mythos itself
Amazon Web Services — the cloud infrastructure backbone for thousands of enterprises and government agencies
JPMorgan Chase — representing the financial sector, which has been among the most alarmed by Mythos’s implications
Nvidia — the chipmaker whose next-generation GPUs powered Mythos’s training

Mythos is the first of a new generation of AI models trained on next-generation GPUs, and its capabilities have sent ripples far beyond Silicon Valley. German banks reported consulting authorities and cyber experts about the risks, while the Bank of England confirmed that AI risk testing had intensified following the announcement.

The Government Angle: Why This Matters for Policy

Anthropic’s decision to restrict Mythos access and partner directly with major institutions is widely interpreted as a strategic move to demonstrate responsible AI development to regulators. The company has faced increasing scrutiny from government bodies concerned about the dual-use nature of advanced AI systems — the same capabilities that can harden defenses can also be weaponized.

By voluntarily withholding Mythos from public access and creating a structured, vetted program for its use, Anthropic is signaling to policymakers that it takes AI safety seriously. This approach could help the company regain favor with government agencies that have been debating how to regulate powerful AI systems without stifling innovation.

The timing is significant. Federal agencies are already evaluating whether to give selected organizations access to Anthropic’s technology for cybersecurity purposes, according to reports from government technology outlets. The White House has reportedly considered providing federal agencies with access to Mythos for defensive testing — a move that would effectively endorse Anthropic’s controlled-release model.

Expert Opinions: Alarm vs. Perspective

Despite the dramatic announcement, the cybersecurity community remains divided on Mythos’s true significance.

Peter Swire, professor at Georgia Tech’s School of Cybersecurity and Privacy and former advisor to the Clinton and Obama administrations, noted that “a large fraction of cybersecurity professors believe this is pretty much what was expected, and pretty much more of the same.” He called Anthropic’s announcement “very dramatic and a PR success, if nothing else.”

Ciaran Martin, professor at Oxford’s Blavatnik School of Government and former CEO of the U.K.’s National Cyber Security Center, echoed that sentiment: “It’s a big deal, but it’s unlikely to prove to be the end of the world. I would not be at the more apocalyptic end of the scale.”

Both experts pointed out important caveats in the testing conditions. During AISI’s evaluation, Mythos faced software environments with near-nonexistent defenses — lacking many protections present in real-world systems. Martin compared this to “a soccer forward scoring a goal against the world’s worst goalkeeper.”

However, neither expert dismisses the threat. Swire warned that “one risk after Mythos is that it will be easier to turn a vulnerability, a known flaw, into an exploit — something that somebody actually takes advantage of.” He advised that “every cybersecurity defender should take Mythos seriously, but the expected harm to defense is likely to be far lower than the worst-case scenarios would suggest.”

The Twenty-Year Equilibrium Just Broke

For roughly two decades, the cybersecurity landscape has operated under a fragile equilibrium: attackers always had an edge in finding zero-day vulnerabilities, but the difficulty and cost of exploitation kept the playing field roughly level. Skilled human hackers were the bottleneck.

Mythos changes that equation. An AI that can find and exploit vulnerabilities at superhuman speed and scale shifts the advantage decisively toward the offensive side — at least until defenders can deploy similar AI tools at scale. The model’s system card contains what some analysts call a “quiet bombshell”: it describes capabilities that allow the model to cover its tracks after exploitation, making detection even harder.

A New Chapter in the AI-Cybersecurity Arms Race

The emergence of Mythos signals the opening of a new chapter in the long-running arms race between offensive and defensive cybersecurity. Historically, each generation of tools — from automated vulnerability scanners to fuzzing frameworks — has been met with corresponding defensive innovations. AI represents a qualitative leap because it compresses the timeline from discovery to exploitation from weeks or months into minutes.

Security vendors are already responding. Companies like CrowdStrike, Palo Alto Networks, and Mandiant have publicly stated that their teams are evaluating AI-powered threat detection capabilities in response to Mythos-class models. The defensive side of the equation is adapting, but the transition requires significant investment in both technology and talent.

The financial sector, which faces some of the highest regulatory scrutiny and the most severe potential consequences from a breach, has been the most reactive. Beyond the German banks’ consultations with authorities, the U.S. Treasury Department and Federal Reserve reportedly warned bank CEOs about the cybersecurity risks posed by Anthropic’s new model in late April 2026. The message from regulators was clear: take this seriously, and take action now.

What Organizations Should Do Now

Whether you view Mythos as a watershed moment or an incremental advance, the practical implications for organizations are clear:

Accelerate patch management. With 99% of Mythos-discovered vulnerabilities still unpatched, the window of exposure is widening. Prioritize critical infrastructure and internet-facing systems.
Assume AI-assisted attacks are imminent. Even if Mythos itself remains restricted, the techniques it demonstrates will be replicated. Update threat models accordingly.
Invest in AI-powered defense. The same capabilities that make Mythos dangerous offensively can be deployed defensively. Organizations that don’t adopt AI-driven security tools will fall behind.
Strengthen detection and response. If vulnerabilities are found faster than ever, the differentiator becomes how quickly you can detect exploitation and respond.
Engage with government initiatives. Federal agencies are developing frameworks for AI cybersecurity collaboration. Early engagement could provide access to shared intelligence and testing resources.

The Broader Picture

Anthropic’s approach with Mythos reveals a new playbook for AI companies navigating the tension between innovation and safety. By announcing capabilities while restricting access, the company generates maximum attention while minimizing immediate risk — and positions itself as a responsible partner for government agencies.

Whether this strategy succeeds in rebuilding Anthropic’s relationship with regulators remains to be seen. But one thing is clear: the era of unrestricted AI model releases may be ending. Mythos could mark the beginning of a new paradigm where the most powerful AI systems are treated not as products to be shipped, but as capabilities to be carefully managed — more like nuclear technology than software.

The cybersecurity community’s divided reaction — between alarm and measured perspective — reflects a deeper uncertainty about how AI will reshape the digital security landscape. What everyone agrees on is this: the old rules no longer apply, and organizations that continue to operate as if they do will find themselves increasingly exposed.

The Bottom Line

Anthropic’s Mythos is not just a technical achievement — it’s a political statement. By choosing to restrict access, partner with select institutions, and engage directly with government concerns, Anthropic is trying to chart a middle path between reckless acceleration and paralysis by caution.

The model’s capabilities are real, the risks are significant, and the appropriate response is neither panic nor complacency but preparedness. Organizations that start treating AI-assisted cyber threats as a present reality — not a future possibility — will be the ones best positioned to weather the changes ahead.

What’s your take on Anthropic’s approach to Mythos? Is controlled access the right model for dangerous AI capabilities, or should these systems never have been built in the first place? Share your thoughts in the comments below.

Anthropic’s New Cybersecurity Model Could Get It Back in the Government’s Good Graces