Claude Mythos meets Claude Reality

Thomas Friedman recently interrupted his geopolitics column concerning the war with Iran in the New York Times to announce a 'stunning advance in artificial intelligence.' Quoting Friedman directly: 'Holy cow, super intelligent AI is arriving faster than anticipated.' He was talking about Anthropic's new unreleased model, Claude Mythos, which supposedly identifies and exploits decades-old zero-day vulnerabilities so efficiently that releasing it would collapse global infrastructure.

A large portion of the internet currently believes Anthropic just invented the WOPR supercomputer from the 1983 Matthew Broderick movie WarGames. Except, they didn't. And Anthropic probably knows it.

The illusion of the new zero-day

The primary narrative Anthropic is pushing is that Mythos displayed a terrifying, emergent ability to find zero-day vulnerabilities no one knew existed. The reality is that security researchers have been using LLMs to find vulnerabilities and program exploits since the beginning of consumer AI.

A 2024 study from IBM titled 'LLM agents can autonomously exploit one day vulnerabilities' found that standard GPT-4 successfully exploited 87 percent of the vulnerabilities it was presented. Even more damning, Anthropic's own release notes for their earlier, less powerful Opus 4.6 model stated their researchers used it to find 'over 500 exploitable zero-day vulnerabilities, some of which are decades old.'

That is almost word-for-word the exact same terminology Anthropic used to describe the supposedly revolutionary Mythos. Our infrastructure survived Opus. It will survive Mythos.

The three-billion-parameter elephant in the room

Anthropic dropped a terrifying press release and formed 'Project Glasswing' to keep Mythos out of the wrong hands, granting access only to a select group of tech giants. It worked perfectly. The internet panicked.

But then independent security researchers actually looked at the flagship zero-days Anthropic bragged about discovering. The CEO of HuggingFace noted that they ran those exact vulnerabilities through cheap, open-weight models. Eight out of eight models detected the flagship FreeBSD exploit Mythos found, including a tiny 3.6-billion-parameter model that costs just 11 cents per million tokens to run. Furthermore, a 5.1-billion-parameter open model recovered the core chain of the 27-year-old OpenBSD bug that Anthropic showcased.

Security researcher Stanislav Fort confirmed the exact same findings, testing the Mythos showcase vulnerabilities and easily replicating the results on tiny, 3-billion-parameter models. As renowned security expert Bruce Schneier summed it up cleanly: 'You don't need mythos to find the vulnerabilities they found.'

The UK AI Security Institute test results

When the UK's AI Security Institute (AISI) ran actual capture-the-flag tests on Mythos, the results painted a picture of slow, steady improvement rather than a massive leap. In a beginner CTF challenge using a 2.5 million token budget, Mythos simply clustered at the top alongside GPT-4 and Opus 4.6. In fact, older models like GPT-4 actually beat it in several metrics.

In the AISI's hardest contrived scenario, which involved a 32-step sequence to break into a loosely protected system, Claude Opus 4.6 made it through 16 steps before getting stuck. Mythos made it through 22 steps. That is a noticeable bump of roughly six steps. It is not the sudden crossing of the Rubicon that demands emergency infrastructure coalitions.

The $60-billion marketing pivot

For two years, Anthropic's CEO Dario Amodei has promised data centers full of 'geniuses' that will automate massive swaths of the economy toward artificial general intelligence. That is the storyline that justifies $60 billion in investor capitalization.

So why, for their biggest, most intensely trained model to date, are they pivoting the narrative to finding bugs in routine C++ code? Finding vulnerabilities is exactly what AI skeptics have always said LLMs would be useful for. It is the boring, nerdy stuff we have known about since GPT-3.

The hyped-up cybersecurity dread is a distraction. If Mythos was actually performing the massive white-collar automation Amodei predicted, they wouldn't need to scare journalists about zero-days. None of the massive leaps in capability between GPT-3 and GPT-4 caused Thomas Friedman to declare that AI was suddenly more important than a global war. Anthropic pushed the big red cyber-fear button because they needed a headline, and they knew the press would blindly run with it.

The ironic coda

There is an incredible irony to all of this. Just a week before Mythos was announced, the source code for Claude Code leaked online. Security researchers immediately found massive security vulnerabilities in it. Anthropic apparently forgot to run their own code through their elite bug-finding AI.

Mythos is a capable model, but the only truly terrifying thing about this release is how easily the tech press bought the ghost story without checking the math. The real story isn't that AI is becoming too dangerous to release. It is that the industry is struggling to deliver the AGI miracles they promised, and they are hoping a ghost story will buy them more time.