Anthropic's Unreleased Claude Mythos Preview Finds 10,000+ Vulnerabilities in One Month
Anthropic's unreleased Claude Mythos Preview model has discovered more than 10,000 vulnerabilities across partner organizations in its first month of deployment through Project Glasswing. The company reports partners are finding bugs at 10x their previous rate, with Cloudflare discovering 2,000 bugs and Mozilla finding 271 Firefox vulnerabilities — 10x more than with previous Claude models.
Anthropic's Unreleased Claude Mythos Preview Finds 10,000+ Vulnerabilities in One Month
Anthropic's unreleased Claude Mythos Preview model has discovered more than 10,000 vulnerabilities across partner organizations in the first month since launching Project Glasswing in April 2026, according to the company's initial report published May 23.
Deployment Results
According to Anthropic, most Project Glasswing partners have "each found hundreds of critical- or high-severity vulnerabilities in their software" using Mythos Preview. The company claims partners' bug-finding rates have increased by more than 10x:
- Cloudflare: 2,000 bugs total, 400 classified as high or critical severity
- Mozilla: 271 vulnerabilities in Firefox, 10x more than found in an older browser version using another Claude model
- Microsoft: Bug discoveries via Mythos Preview are reportedly why the company announced patch releases will "continue trending larger for some time"
Anthropic also scanned 1,000 open-source projects over recent months using Mythos Preview, identifying 6,202 high- and critical-severity vulnerabilities out of 23,019 total findings.
A security research firm recently claimed it used Mythos' bug-finding capabilities to breach macOS, though Anthropic did not include this in its official report.
Why Mythos Preview Remains Unreleased
Anthropic states it has not released Mythos Preview publicly because "no company (including itself) has developed safeguards strong enough to prevent models like it from being misused." The company intends to release "Mythos-class models" when adequate safeguards become available.
Current Project Glasswing partners include Amazon Web Services, Apple, CrowdStrike, Google, JPMorgan Chase, NVIDIA, Palo Alto Networks, Cloudflare, Mozilla, and Microsoft. Anthropic plans to expand availability to US and other government partners.
Financial Context
According to The Wall Street Journal, Anthropic is on track to post $10.9 billion revenue with $559 million operating profit for the quarter ending June 2026, marking potential first-time profitability since its 2021 founding. The company does not expect sustained profitability in subsequent quarters due to planned investments in computing resources.
What This Means
Mythos Preview represents a capability level that Anthropic considers too dangerous for public release — a rare acknowledgment that a model's offensive security capabilities outpace available defensive safeguards. The 10x improvement in partner bug-finding rates, if sustained, could significantly alter vulnerability disclosure timelines and patch cycles across the software industry. The government partnership expansion suggests cybersecurity applications may drive near-term AI model restrictions more than misinformation or other commonly cited risks.
Related Articles
Anthropic Python SDK v0.104.0 adds thinking token count estimates for streaming responses
Anthropic released version 0.104.0 of its Python SDK on May 21, 2026. The update adds support for a thinking-token-count beta feature that provides estimated token counts in thinking block deltas when streaming responses from reasoning models.
Google releases Gemini Omni Flash video generation model with conversational editing, withholds speech synthesis
Google DeepMind released Gemini Omni Flash, the first model in its new Omni family that generates and edits video from image, audio, video, and text inputs. The model is rolling out to Gemini app subscribers and YouTube Shorts with a 10-second clip limit, while speech-editing capabilities remain withheld pending safety testing.
Tencent Releases Hy-MT2 Translation Models: 1.8B, 7B, and 30B-A3B Support 33 Languages
Tencent released Hy-MT2, a family of multilingual translation models available in 1.8B, 7B, and 30B-A3B (MoE) sizes. All models support translation among 33 languages and follow translation instructions in multiple languages. The 1.8B model can be compressed to 440MB using 1.25-bit AngelSlim quantization.
NVIDIA releases Nemotron-Labs-Diffusion-14B with tri-mode decoding achieving 3.3x speed-up on GB200
NVIDIA released Nemotron-Labs-Diffusion-14B, a 14-billion parameter language model that supports three decoding modes by switching attention patterns during inference. The model achieves 850 tokens per second on GB200 hardware at concurrency 1, representing a 3.3x speed-up over standard autoregressive decoding and outperforming Qwen3-8B-Eagle3 by 2.2x in self-speculation mode.
Comments
Loading...