Anthropic leaked its scariest model

Anthropic had a bad weekend. Roughly 3,000 unpublished internal documents spilled onto the open web after a misconfigured CMS left assets public by default—no authentication required. Among them: a draft describing Claude Mythos, codenamed Capybara, as “the most capable we’ve built to date” and “a step change” above Opus 4.6, with dramatically higher scores in software coding, academic reasoning, and cybersecurity tasks. Anthropic’s own language is the uncomfortable part—the document warns the model is “currently far ahead of any other AI model in cyber capabilities” and could “exploit vulnerabilities in ways that far outpace the efforts of defenders.” CrowdStrike fell 7% and Palo Alto Networks dropped 6% on the news. Anthropic confirmed it’s testing Mythos with a small group of early access customers, and attributed the leak to human error. The company that built its brand on AI safety just accidentally published evidence that its newest model may be the most dangerous one in existence.

The models keep getting more capable, and the people running the economy have noticed. Fed Chair Jerome Powell told Harvard students today that “these large language models make people much more productive”—and added that AI is making him personally more productive, too. Powell urged undergrads to “invest the time to really master” the technology, while acknowledging it’s “a challenging time to enter the labor market.” He also noted that the AI data center boom is “actually probably pushing inflation up” in the short term, and said he’d observed “meaningfully higher productivity” over several years. The capital markets are responding accordingly: Harvey, the legal AI platform, raised $200 million at an $11 billion valuation—up from $8 billion in December—co-led by Sequoia and GIC, Singapore’s sovereign wealth fund. Sequoia has now led three of Harvey’s funding rounds, and the platform serves 100,000+ lawyers across 60 countries, including 50+ AmLaw 100 firms. When the Fed chair tells students to learn AI and a legal AI startup jumps $3 billion in valuation in three months, the productivity thesis isn’t theoretical anymore—it’s repricing entire industries.

And those productivity gains are moving from text generation to direct computer control. Claude’s Computer Use API—the tool that turns any GUI into a programmable endpoint—now scores 72.5% on OSWorld, against a 72.4% human baseline, at roughly $0.08–$0.15 per 20-step task on Sonnet 4.6 pricing. Each session carries about 2,700 tokens of fixed overhead, and every screenshot at 1,024×768 costs another 1,048 tokens. We walked through the full implementation—sampling loop, Docker setup, coordinate scaling, and security hardening—in our latest tutorial. Compare that per-task cost to UiPath’s unattended licenses at roughly $420 per month, and the economics of desktop automation start to look very different. The same company that accidentally leaked a model it describes as unmatched in cyber capabilities is also shipping the tools that let AI agents click through your desktop autonomously—and right now, the market is pricing that as progress.

Latest from PulseMark

How to Build a Desktop AI Agent with Claude Computer Use API The sampling loop architecture, coordinate scaling pitfalls, Docker sandboxing, and a cost comparison that puts UiPath’s $420/month licenses against Sonnet 4.6’s $0.08–$0.15 per task. Read the tutorial →