Anthropic unveils the Glasswing Initiative, with Apple, Microsoft, Google, and others queueing up to scan their own code using the undisclosed Claude Mythos model.
According to 1M AI News monitoring, Anthropic announced its Project Glasswing cyber security initiative on Tuesday, along with the official unveiling of the previously leaked new model Claude Mythos Preview. This model was not publicly released for sale as Anthropic believes its network defense capabilities are too strong and wants the defense side to use it first. Apple, Microsoft, Google, AWS, NVIDIA, and 11 other institutions as founding partners will use Mythos Preview to scan their systems for security vulnerabilities, while over 40 key software infrastructure institutions have also been granted access.
Anthropic CEO Dario Amodei said in a release video, "We didn't specifically train it for cybersecurity, we trained it for programming, but the side effect of being good at programming is being good at cybersecurity. More powerful models will continue to emerge from us and other companies, so we need a response plan."
In the past few weeks, Mythos Preview has discovered thousands of high-risk zero-day vulnerabilities in all major operating systems and browsers, mostly autonomously and without human guidance. Three cases that have already been patched include:
1. A 27-year-old vulnerability in OpenBSD (known for its security and widely used in firewall operating systems) that allows an attacker to remotely crash the target machine as soon as they connect
2. A 16-year-old vulnerability in FFmpeg (an open-source library widely used for video encoding and decoding) that automated testing tools failed to detect even after running that line of code 5 million times
3. Chained exploits of multiple vulnerabilities in the Linux kernel that can elevate privileges from a regular user to full control of the entire machine
Full list of founding partners: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorgan Chase, Linux Foundation, Microsoft, NVIDIA, Palo Alto Networks. Anthropic has allocated up to $100 million in model usage credits, and has also donated $4 million to open-source security organizations ($2.5 million to Alpha-Omega and OpenSSF under the Linux Foundation, and $1.5 million to the Apache Software Foundation).
In benchmark tests, Mythos Preview has significantly widened the gap with the previous flagship Claude Opus 4.6:
1. CyberGym Network Security Vulnerability Reproduction: 83.1% vs 66.6%
2. SWE-bench Verified: 93.9% vs 80.8%
3. SWE-bench Pro: 77.8% vs 53.4%
4. GPQA Diamond: 94.6% vs 91.3%
5. Humanity's Last Exam (including tools): 64.7% vs 53.1%
Once the $100 million credits are used up, partners will pay $25 per million tokens input and $125 per million tokens output, supporting the Claude API, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Anthropic plans to first deploy security protection on the next version of Claude Opus as groundwork for opening up Mythos-level models in the future.
Anthropic's Head of Advanced Red Team, Logan Graham, told Wired, "This capability will become mainstream within 6 to 24 months, and many of the assumptions our current security systems rely on may fail."
On the same day, Anthropic disclosed that its annualized revenue has exceeded $30 billion (around $9 billion by the end of 2025), and announced agreements for several gigawatts of computing power with Google and Broadcom, with reports suggesting an IPO as early as October of this year.
