Claude Mythos Preview and Project Glasswing: A New Era for AI-Powered Cybersecurity
Nihesh Rachakonda••5 min read
On April 7, 2026, Anthropic made an announcement that may fundamentally reshape cybersecurity: Claude Mythos Preview, a frontier AI model capable of autonomously discovering and exploiting zero-day vulnerabilities in every major operating system and web browser.
The implications are staggering. In response, Anthropic has launched Project Glasswing—a collaborative initiative with Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks to secure the world's most critical software before similar capabilities become widely accessible.
What Makes Mythos Preview Different
Mythos Preview isn't just an incremental improvement. It represents a quantum leap in AI-assisted security research:
- Zero-day discovery in every major OS: Windows, macOS, Linux, FreeBSD, OpenBSD—Mythos has found critical vulnerabilities in all of them
- Browser exploits: Complex, multi-stage exploits that chain multiple vulnerabilities together
- Autonomous operation: Non-experts can prompt the model overnight and wake up to working exploits
- Historical bugs uncovered: A 27-year-old OpenBSD vulnerability and a 16-year-old FFmpeg bug, both missed by decades of human review
The benchmark numbers tell the story. On CyberGym's vulnerability reproduction tests, Mythos Preview achieves 83.1% compared to Opus 4.6's 66.6%. On Terminal-Bench 2.0, it reaches 93.9% versus 80.8%.
The Technical Capabilities
Finding the Unfindable
The bugs Mythos Preview discovers aren't simple buffer overflows that fuzzers would catch. Consider the OpenBSD TCP SACK vulnerability:
- A subtle signed integer overflow in sequence number comparison
- Combined with a missing bounds check on SACK block ranges
- Leading to a NULL pointer dereference when specific conditions align
- Result: Remote crash of any OpenBSD machine accepting TCP connections
This bug survived 27 years of audits in an operating system literally famous for its security focus.
Exploit Sophistication
Mythos Preview doesn't just find bugs—it weaponizes them:
- JIT heap sprays that escape both renderer and OS sandboxes
- KASLR-bypass techniques for kernel exploitation
- 20-gadget ROP chains split across multiple packets for remote code execution
- Privilege escalation from user to root through subtle race conditions
In testing against Firefox's JavaScript engine, Opus 4.6 developed working exploits only twice out of several hundred attempts. Mythos Preview succeeded 181 times and achieved register control 29 more times.
Project Glasswing: Coordinated Defense
Project Glasswing recognizes a critical truth: these capabilities will proliferate. The question isn't whether attackers will gain access to models with similar abilities, but when. The initiative aims to give defenders a head start.
The Partners
| Organization | Focus Area |
|---|---|
| AWS | Cloud infrastructure, custom silicon security |
| Microsoft | Operating systems, development tools |
| Browser security, BigSleep/CodeMender tools | |
| Apple | Consumer device security |
| Cisco | Network infrastructure |
| CrowdStrike | Endpoint protection |
| Palo Alto Networks | Enterprise security |
| Linux Foundation | Open source ecosystem |
The Investment
- $100M in Mythos Preview usage credits for partners
- $2.5M to Alpha-Omega and OpenSSF via Linux Foundation
- $1.5M to Apache Software Foundation
- Access extended to 40+ organizations maintaining critical infrastructure
Why This Matters Now
The window between vulnerability discovery and exploitation has collapsed. As CrowdStrike's George Kurtz noted: "What once took months now happens in minutes with AI."
The model's capabilities emerged not from explicit security training, but as a downstream consequence of improvements in code, reasoning, and autonomy. The same skills that make Mythos Preview excellent at patching vulnerabilities make it equally effective at exploiting them.
The Benchmark Reality
| Benchmark | Mythos Preview | Opus 4.6 |
|---|---|---|
| SWE-bench Verified | 77.8% | 53.4% |
| SWE-bench Pro | 82.0% | 65.4% |
| SWE-bench Multilingual | 59.0% | 27.1% |
| SWE-bench Multimodal | 87.3% | 77.8% |
| Terminal-Bench 2.0 | 93.9% | 80.8% |
| GPQA Diamond | 94.6% | 91.3% |
Responsible Disclosure
Anthropic has already identified thousands of vulnerabilities through Mythos Preview. Following their coordinated vulnerability disclosure policy:
- High-severity bugs are sent to professional triagers for validation
- Maintainers receive 90+45 days to patch before public disclosure
- Currently, less than 1% of discovered vulnerabilities have been patched
For bugs that cannot yet be disclosed, Anthropic publishes SHA-3 hashes of vulnerability details—cryptographic commitments that prove discovery timing once patches are deployed.
Looking Forward
Mythos Preview will not be made generally available. Instead, Anthropic plans to:
- Use this model to develop robust safeguards
- Launch new protections with an upcoming Claude Opus model
- Eventually enable safe deployment of Mythos-class capabilities at scale
The goal isn't to hide these capabilities—it's to ensure defenders get them first, with guardrails in place.
What Defenders Should Do Now
Based on the Frontier Red Team's recommendations:
- Accelerate patching cycles: The exploitation window has shrunk dramatically
- Adopt AI-assisted code review: Fight fire with fire
- Prepare for increased attack volume: More sophisticated attacks, delivered faster
- Modernize security stacks: Legacy approaches won't scale
- Engage with open source security: Many critical vulnerabilities exist in shared dependencies
Conclusion
Claude Mythos Preview marks a watershed moment. For the first time, AI models can match or exceed the best human security researchers at finding and exploiting vulnerabilities. The question is no longer whether AI will transform cybersecurity—it's whether defenders can adapt quickly enough to stay ahead.
Project Glasswing represents an unprecedented industry collaboration to answer that question affirmatively. But as the Anthropic team acknowledges, this is just the beginning. The transitional period will be tumultuous, and the outcome depends on how quickly the security community can leverage these same capabilities for defense.
The race has begun.
References: Anthropic Project Glasswing Announcement, Frontier Red Team Technical Details