.png){kind=logo}
Anthropic’s Mythos artificial intelligence model found vulnerabilities in classified US government systems within hours during a security test, according to a US official. The disclosure has raised questions about government networks, open-source software and crypto systems as AI tools gain stronger cyber capabilities.
The official stressed that finding a weakness did not mean Mythos exploited it in the same period. That distinction casts doubt on claims that the model ‘broke into’ classified systems within hours. The tests formed part of Project Glasswing, a program involving Anthropic and US intelligence agencies.
Senator Mark Warner referred to the tests during a June 11 congressional hearing. Citing NSA and US Cyber Command chief General Joshua Rudd, Warner said, “This tool broke into almost all of our classified systems, not in weeks, but in hours.”
Still, the wording is disputed. A US official said Mythos identified flaws quickly but did not necessarily use them to gain access. Anthropic and the NSA declined to provide more details, leaving the scale and nature of the findings unclear.
Project Glasswing aims to find and patch software flaws before hostile groups can use them. Mythos has also found weaknesses in common software. Reports say it detected a 27-year-old flaw in OpenBSD and a 16-year-old bug in the FFmpeg video library.
The model has also drawn attention in the crypto sector. In early June, Mythos reportedly found a critical flaw in Zcash’s Oraudits,shielded pool. The bug had stayed undetected for four years before developers moved to patch the system.
That case has widened the debate over AI and blockchain security. Crypto networks often rely on open-source code, public audits and long review periods. However, Mythos has shown that old code can still contain hidden weaknesses after repeated checks.
Bitcoin has faced years of testing, while developers have made few changes to its core system. Still, security experts cannot assume that age alone makes any network safe. AI models can now test code faster than smaller audit teams.
Reports show that crypto hacks have risen sharply in the second quarter of 2026. AI tools may help defenders find flaws earlier, but the same tools can help attackers study code, build exploits and automate parts of an attack.
The US government has taken two different approaches to Anthropic’s models. Intelligence agencies have used Mythos for defensive testing. Meanwhile, the Commerce Department ordered Anthropic on June 12 to block foreign nationals from using Mythos 5 and Fable 5.
Anthropic complied by disabling access for customers, although it questioned whether the directive matched the stated security concern. More than 100 cybersecurity experts and executives later urged the government to reverse the measure.
They argued that Mythos is ‘quite good’ at finding and weaponizing flaws but ‘not uniquely good at these tasks.’ Their letter warned that limiting defensive access could weaken US security while rival states develop similar tools.
The White House also blocked Anthropic’s plan to expand Mythos access from about 50 organizations to 120. Officials cited national security concerns and limits on Anthropic’s computing capacity. Access stays narrow as agencies assess how to use the model without widening the risk.
Also Read: Claude Code vs Codex: How to Use Both Effectively
