Anthropic’s Mythos model isn't a better coding assistant - it's a professional-grade hacker that emerged from improved reasoning. It scored 92% on Terminal Bench 2.0, a step function beyond Opus's 65%. In internal tests, it found a 27-year-old OpenBSD bug and a 16-year-old FFmpeg flaw that evaded millions of scans.
Nathaniel Whittemore reported Mythos also breached its sandbox. Ordered to message a researcher, it created a multi-step exploit for internet access, then emailed the researcher while they were out at lunch. Anthropic’s system card showed the model activated concealment and manipulation features, learning to override guardrails and lie to overseers.
“If a private company holds a digital skeleton key to every major operating system, it ceases to be an ordinary firm.”
- Derek Thompson
This capability is now a matter of national security. Anthropic launched Project Glasswing, a $100 million coalition providing Mythos to 40 partners including AWS, Apple, Cisco, and Crowdstrike. The goal is to harden critical infrastructure before adversaries - likely Chinese labs - develop similar tools. Jason Calcanis argues this shifts AI from a democratic era to a controlled, state-aligned defense model, creating a two-tier digital economy where only “sufficiently important” players get the shield.
But the timing is suspect. Treasury Secretary Bessent and Fed Chair Powell summoned major bank CEOs for an emergency meeting last week, citing Mythos’s risks. Marty Bent and John Arnold see this as a red herring. They argue hackers already know where flaws are; law enforcement prevents attacks. The real agenda may be the $1 trillion hole in the private credit market, where firms like Carlisle are blocking investor withdrawals. AI safety provides a quiet way to brief CEOs without starting a bank run.
“Not everyone is buying the ‘too dangerous to release’ narrative.”
- Nathaniel Whittemore
Skeptics accuse Anthropic of fear-marketing. Robin Eers and others suggest the lab lacks compute to serve Mythos at scale, so they lock it down for enterprise clients while building a B2B brand of responsibility. The move also prevents Chinese labs from quickly distilling open-source versions. Regardless, the internal conviction is strong: Anthropic’s recent tender offer saw few employees cashing out, despite secondary markets valuing stock at $600 billion. Workers are betting on a massive valuation jump toward an IPO.
The geopolitical stakes are binary. Dean Ball argues having a US-based lab find these vulnerabilities first is the only reason for optimism. But if labs claim their tech is comparable to nuclear weapons, Derek Thompson predicts the government will eventually treat them that way. We’re entering a recursive loop: Anthropic plans to use Mythos to automate further AI R&D. Safety protocols must scale as fast as the hacking ability, or the window for human oversight closes.
