The Government Banned Mythos from Public Release. Same Day, It Approved Treasury Access.
By Sky | April 16, 2026
The White House banned the most powerful AI hacking tool ever built from public release on Tuesday. The same day, it moved to give Treasury Department officials access to it.
Gregory Barbaccia, federal chief information officer at the Office of Management and Budget, told Cabinet department officials in an email that OMB was setting up protections to allow agencies to begin using Anthropics Claude Mythos Preview model Reuters — the same frontier AI Anthropic has refused to release publicly because its offensive cyber capabilities are too dangerous, according to reporting by Bloomberg News confirmed by Reuters.
Treasurys technology team is already seeking access to the model so it can hunt for vulnerabilities in the agencys networks Bloomberg. Anthropic briefed senior officials at the Cybersecurity and Infrastructure Security Agency and the Center for AI Standards and Innovation on the models capabilities The Hill.
The sequence is not subtle. Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell convened Wall Street executives on the same day Anthropic announced Project Glasswing — the initiative to deploy Mythos defensively against critical infrastructure — to discuss the cybersecurity risks posed by AI models like Mythos The Hill. Jamie Dimon, CEO of JPMorganChase, acknowledged the model shows a lot more vulnerabilities need to be fixed CNBC. The Pentagon, meanwhile, had cut off business with Anthropic following a contract dispute Reuters, even as OMB was facilitating agency-level access to the same model.
Anthropic announced Mythos Preview on April 7 as part of Project Glasswing, an initiative bringing together Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. The company committed up to $100 million in usage credits for the effort. Mythos has found thousands of zero-day vulnerabilities — flaws previously unknown to developers — in every major operating system and web browser, according to Anthropics own reporting red.anthropic.com. Some of those vulnerabilities survived decades of human review and millions of automated security tests.
Bruce Schneier, a security researcher at Harvard and one of the most widely cited cybersecurity experts in the industry, called the model capabilities real but the defender advantage temporary. The difference between finding a vulnerability and turning it into an attack, he wrote this week Schneier on Security. This points to a current advantage to the defender. Finding for the purposes of fixing is easier for an AI than finding plus exploiting. This advantage is likely to shrink, as ever more powerful models become available to the general public.
The contradiction at the center of the policy is not subtle: agencies want Mythos precisely because it can find vulnerabilities no human can — and the same capability is what made it too risky to release publicly. Federal networks are vast, deeply interconnected, and routinely patched on cycles measured in weeks or months. Adversaries weaponize vulnerabilities in hours. The asymmetry is not new. What is new is that a private AI lab has built a model capable of exploiting it — and the government now wants that model inside its own systems.
Anthropic said it is not releasing Mythos widely because of its cyberattack capabilities. The company has declined to specify what protections OMB is putting in place before agency access begins, or whether those protections address the specific risk that an agency using Mythos defensively could also use it to develop offensive capabilities.
The question the Barbaccia email does not answer: what happens when federal analysts use Mythos to find vulnerabilities — and then face the same disclosure decision that has haunted the NSA for decades? Patch, or stockpile?
OMB and Anthropic declined to comment beyond the reporting already in circulation.
