Anthropic's Claude Mythos can find zero-day exploits faster than defenders can patch them
Anthropic announced Claude Mythos Preview, a new frontier model with advanced reasoning capabilities that can identify and chain together multiple vulnerabilities into novel attacks—abilities the company says outpace current defensive capabilities. The model has already discovered thousands of high-severity vulnerabilities including a 27-year-old OpenBSD flaw and exploits for multiple operating systems. To manage the risk, Anthropic launched Project Glasswing, granting early access to 40+ companies including Apple, Google, Microsoft, and Cisco, providing $100M in usage credits for defensive security work.
Claude Mythos Preview at a glance
How Claude Mythos Preview ranks now
| Benchmark | Score | Rank | Source |
|---|---|---|---|
| GPQA | 94.6% | 6 of 165 | the-decoder.com ↗ |
| SWE-bench Verified | 93.9% | 3 of 80 | platformer.news, Apr 2026 ↗ |
Rank is against every model in our database with a sourced score on that benchmark, each counted once with its best score.
Anthropic's Claude Mythos Can Find Zero-Day Exploits Faster Than Defenders Can Patch Them
AnthropIC announced Claude Mythos Preview on Tuesday after accidentally leaking its existence two weeks prior—a frontier AI model the company explicitly states poses serious new cybersecurity risks that outpace current defensive capabilities.
The model represents a significant capability leap. According to Anthropic's announcement, Mythos achieves 93.9% on SWE-bench Verified, a 13-percentage-point improvement over Claude Opus 4.6's 80.8% score. This performance gain comes directly from improvements in reasoning—the same general capability improvements every AI lab is pursuing—rather than specialized cyber training.
What Mythos Can Already Do
Mythos has identified thousands of high-severity vulnerabilities across major operating systems and web browsers, including:
- A vulnerability in OpenBSD that evaded detection for 27 years
- A flaw in FFmpeg video encoder that survived 5 million automated tests
- Multiple Linux kernel vulnerabilities that could enable complete machine compromise
Crucially, the model can chain together separate vulnerabilities into novel attacks—a capability current models lack. Combined with AI systems' growing ability to operate without human supervision for extended periods, researchers say this represents an inflection point in cybersecurity risk.
Project Glasswing: Controlled Deployment
Rather than releasing Mythos broadly, Anthropic launched Project Glasswing, a coalition of 40+ companies including Apple, Google, Microsoft, Cisco, and Broadcom. Participants receive $100 million in model usage credits to scan and patch vulnerabilities in their own systems and critical open-source infrastructure. Anthropic is also donating $4 million to open-source security initiatives.
Alex Stamos, chief product officer at Corridor and former security lead at Facebook and Yahoo, called Glasswing "a big deal, and really necessary." He warned that "open-weight models will catch up to foundation models in bug finding within six months, at which point every ransomware actor will be able to find and weaponize bugs without leaving traces."
Cisco's chief security officer Anthony Grieco stated: "AI capabilities have crossed a threshold that fundamentally changes the urgency required to protect critical infrastructure from cyber threats, and there is no going back."
The Uncomfortable Premise
Glasswing rests on a deeply uncomfortable foundation: the only way to defend against dangerous AI capabilities is for a safety-focused lab to build them first. This centralizes both power and risk. Anthropic now possesses zero-day exploits for virtually all major software systems—a capability whose theft would represent a severe security threat.
The timing is awkward. The Trump administration is resisting AI regulation while the US government previously attempted to declare Anthropic a supply chain risk after it refused to include mass domestic surveillance and autonomous weapons in Pentagon contracts. Anthropic briefed senior government officials including CISA before the announcement but faces unclear government interest in collaboration.
Timeline Uncertainty
Stamos offered two scenarios: an optimistic timeline where superhuman capabilities find a finite, patchable set of flaws, and a pessimistic one where each model release discovers new classes of vulnerabilities we "never even imagined." The actual outcome remains unknowable.
What this means
Mythos validates Anthropic's founding thesis—that a safety-focused lab building frontier models could discover dangerous capabilities first and lead mitigation efforts. However, it also demonstrates that unregulated AI development can create genuine national security risks when capability gains in general reasoning translate directly into cybersecurity threats. The race between Anthropic's responsible disclosure through Glasswing and the inevitable proliferation of these capabilities to open-weight models may determine whether critical infrastructure survives the next two years of AI progress intact.
Related Articles
Anthropic mathematician Levent Alpöge claims Claude-assisted proof of the Riemann hypothesis, posts it in Lean
Levent Alpöge, a mathematician employed by Anthropic, says he has proven the Riemann hypothesis, first proposed in 1859, working with Claude. He posted the proof coded in Lean, and experts have yet to confirm that it is correct.
Anthropic Adds Experimental Code Mode to Claude Code v2.1.323, Chaining Tool Calls in JavaScript
Anthropic has added an experimental Code Mode to Claude Code v2.1.323. Claude can write JavaScript that chains Bash, Read and Edit calls in one step, and keep talking while long commands run.
Thinking Machines Lab releases Inkling 2 Mini: 26B-parameter looped MoE with 3B active and 1M-token context
Thinking Machines Lab has released Inkling 2 Mini, a multimodal Mixture-of-Experts model with about 26B physical total parameters and 3B active per token. It repeats its middle layers three times, supports a 1M-token context, and ships with a DSpark drafter. Hugging Face reports it leads 32 of 48 benchmarks against four comparable models.
NVIDIA releases Nemotron-Labs-Sigma: 14 open 3B and 8B continuous diffusion language model checkpoints
NVIDIA released Nemotron-Labs-Sigma, a set of 14 open continuous diffusion language model checkpoints at 3B and 8B parameters. According to MarkTechPost, Sigma denoises text in a learned 16-D embedding space, distills to few-step samplers, and decodes up to 4.20x faster than autoregressive decoding.
Comments
Loading...