AI system breaks into secret government networks, Pentagon official says
A secure defense is only as strong as its fastest attacker, and the clock just ran out. A highly restricted artificial intelligence model recently penetrated some of the most secure systems in the United States government within hours. This jaw-dropping revelation has sent shockwaves through the halls of the Pentagon.
The test wasn’t a malicious external hack but an authorized exercise designed to identify hidden vulnerabilities. A U.S. official confirmed that the model exposed these critical flaws during a red-team evaluation. This exercise has completely upended traditional expectations of military cybersecurity.
The tool responsible is known as Mythos, a highly classified unreleased model created by Anthropic. During the assessment, the model systematically analyzed replicated defense environments under a restricted program called Project Glasswing. Federal agencies are now grappling with an offensive tool that operates at superhuman speed.
Inside the high-stakes test of Project Glasswing
Under Project Glasswing, tech companies and intelligence agencies paired up to stress-test critical software. The project initially offered limited access to approximately 50 elite organizations. However, the model’s sheer strength quickly turned a standard security audit into a national security crisis.
During a briefing, Senator Mark Warner expressed sheer astonishment at how quickly the model operated. Citing General Joshua Rudd, the head of the National Security Agency and U.S. Cyber Command, Warner shared a startling update. Warner reported that the tool broke into almost all of these classified systems in hours rather than weeks.
While finding vulnerabilities isn’t the same as actively exploiting them, the time compression is terrifying. Standard corporate and military networks are built on the assumption that security patches can take weeks to deploy. With automated AI agents in play, that security buffer has completely vanished.
Mind-blowing speed and a 27-year-old blind spot
The capabilities demonstrated by Mythos aren’t just minor upgrades; they represent a major shift. In one test, the model flagrantly cracked an elusive bug in OpenBSD that had remained hidden since 1997. This 27-year-old vulnerability had easily survived decades of human auditing and millions of automated tests.
The AI didn’t need special hints or hand-holding to spot this deep kernel flaw. Instead, it methodically swept the code and flagged the relocatable object format bug within minutes. This shows how automated systems ignore human assumptions about which parts of a codebase are stable.
Further testing showed that Mythos autonomously discovered and exploited a 17-year-old remote code execution flaw in FreeBSD. The model successfully analyzed the system’s network file server and executed a functional exploit. Independent teams have already reproduced some of these findings using publicly available models.
The geopolitical fallout and the sudden export ban
This immense power quickly triggered a major confrontation between Anthropic and the White House. On June 12, 2026, the Trump administration issued a sudden directive ordering the export ban of both Mythos 5 and Fable 5. Federal officials classified these frontier models as cyber weapons due to their extreme offensive potential.
This restriction came only 10 days after a new voluntary framework was established to vet advanced AI models. Rather than complying selectively, Anthropic took the extreme step of disabling the models globally. The company explained that verifying the nationality of every individual user in real-time simply wasn’t possible.
The underlying friction goes way deeper than a simple compliance disagreement. Anthropic has steadfastly refused to let the military use its AI for mass surveillance or fully autonomous lethal systems. Its strict usage policy doesn’t permit mass surveillance or lethal actions without human oversight.
Former OpenAI head of robotics Caitlin Kalinowski heavily criticized government-military partnerships. Kalinowski stated that “surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation.” Meanwhile, the Pentagon’s been pushing hard for unrestricted control over these advanced systems.
The Pentagon’s Chief Information Officer, Kirsten Davies, took a firm public stance on national security. Davies declared that some things are just more important than corporate revenue cycles and pre-IPO valuations. Even so, about 100 cybersecurity executives didn’t hold back, signing an open letter warning against these government restrictions.
Industry leaders from Nvidia, Adobe, and Zoom argue that blocking these models severely weakens domestic defense. They pointed out that adversaries like China aren’t sitting on their hands, rapidly developing tools like the Tulongfeng model. Sidelining domestic tools doesn’t stop adversaries; it only hands them a massive strategic advantage.
The bottom line for secure operations
The Project Glasswing test has definitively proven that AI has fundamentally changed the rules of cyber warfare. With models capable of identifying decades-old vulnerabilities in hours, traditional defensive strategies are now obsolete. Security teams must transition to automated, AI-driven defense mechanisms immediately.
Attempting to regulate these technologies through sweeping export bans may actually hamper domestic defensive readiness. If domestic engineers lose access to advanced audit tools, foreign adversaries won’t take long to outpace them. The future of national security depends entirely on mastering these tools rather than hiding from them.
Disclaimer – This list is solely the author’s opinion based on research and publicly available information. It is not intended to be professional advice.
Like our content? Be sure to follow us
