Claude Mythos finds 10,000+ critical CVEs as Anthropic reportedly raises $30B at $900B valuation

Anthropic's Project Glasswing, built on Claude Mythos Preview, surfaced 10,000+ critical CVEs in weeks across thousands of open-source repos using agentic harnesses — a scale shift that reframes vulnerability research from human-led to AI-led. The Register reports Mythos-class models will be released to more users including government customers, though the underlying flaw-finder remains gated while Anthropic finalizes guardrails. A separate dev.to deep-dive includes runnable code for deploying similar AI security pipelines, suggesting the technique is rapidly democratizing.

The capability comes with a sharp edge: UK AISI red-team results show Mythos can fully take over a corporate network in 6 of 10 attempts versus 3 of 10 for GPT-5.5. UK AI Minister Narayan called the gap evidence that 'cyber capabilities are advancing much faster than expected.' OpenAI separately posted a $295K–$445K Preparedness role explicitly to address 'recursive self-improvement' risk, indicating frontier labs view the offensive-cyber curve as a near-term safety concern, not a hypothetical.

The business backdrop is just as striking. Anthropic is reportedly raising $30B at a $900B valuation — a number that, combined with imminent SpaceX, OpenAI, and Anthropic IPO filings flagged by the FT, will test whether public markets accept current AI economics. Microsoft reportedly pulled internal Claude code licenses over Anthropic's pricing despite engineers preferring Claude to GitHub Copilot, and Business Insider reports Claude has 'already won' the AI coding wars inside startups.

What to watch: whether governments accepting Mythos access will accelerate offensive-security disclosure norms, whether the $900B valuation survives diligence, and whether the pricing tension with Microsoft becomes a structural rift between Anthropic and its largest enterprise distribution channel.