AI Cybersecurity After Mythos: Small Open-Weights Models Match Performance of Large-Scale Systems
Following Anthropic's announcement of Claude Mythos Preview and Project Glasswing, new testing reveals that small, affordable open-weights models can recover much of the same vulnerability analysis as high-end systems. While Anthropic's Mythos demonstrated sophisticated capabilities—including finding a 27-year-old OpenBSD bug and creating complex Linux kernel exploits—research suggests that AI cybersecurity capability does not scale smoothly with model size. Instead, the true competitive 'moat' lies in the specialized systems and security expertise built around the models rather than the models themselves. This discovery highlights a 'jagged frontier' in AI development, where smaller models are proving surprisingly effective at identifying zero-day vulnerabilities previously thought to require massive, limited-access AI infrastructure.
Key Takeaways
- Model Size vs. Capability: AI cybersecurity performance is 'jagged' and does not scale linearly with model size; small open-weights models can replicate many findings of larger models.
- The Mythos Benchmark: Anthropic's Mythos autonomously identified thousands of zero-day vulnerabilities, including decades-old bugs in OpenBSD and FFmpeg.
- System-Centric Security: The true advantage in AI security lies in the integrated system and deep expertise rather than the underlying model alone.
- Project Glasswing: A $104M initiative involving usage credits and donations to open-source security organizations to patch critical software.
In-Depth Analysis
The Mythos Announcement and Project Glasswing
On April 7, 2026, Anthropic introduced Claude Mythos Preview and Project Glasswing, a consortium aimed at utilizing limited-access AI to secure critical software infrastructure. Anthropic has committed $100 million in usage credits and $4 million in direct donations to open-source security entities. The technical capabilities showcased were significant: Mythos reportedly discovered thousands of zero-day vulnerabilities across major operating systems and browsers. Notable successes included identifying a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg, alongside constructing sophisticated multi-vulnerability privilege escalation chains in the Linux kernel.
The Jagged Frontier of AI Capabilities
Despite the high-profile nature of Mythos, subsequent testing by researchers like Stanislav Fort indicates that the 'moat' protecting these large models may be thinner than expected. By isolating the code for vulnerabilities showcased by Anthropic and running them through small, cheap, open-weights models, researchers found that these smaller models could recover much of the same analysis. This suggests that AI cybersecurity capability is 'jagged'—it does not improve in a smooth, predictable curve as models get larger. Consequently, the value of an AI security solution is determined more by the system architecture and the security expertise built into it than by the raw scale of the model.
Industry Impact
The findings suggest a shift in the AI security landscape. If small, open-weights models can perform high-level vulnerability analysis, the barrier to entry for both defensive and offensive cybersecurity tools may lower significantly. This democratizes access to advanced security auditing but also emphasizes that the industry's competitive edge will shift toward system-level integration and specialized domain knowledge. Anthropic's massive investment via Project Glasswing validates the importance of AI in open-source security, yet the effectiveness of smaller models suggests that the future of AI-driven security may be more decentralized than previously anticipated.
Frequently Asked Questions
Question: What is Project Glasswing?
Project Glasswing is a consortium of technology companies formed to use Anthropic's Mythos model to find and patch security vulnerabilities in critical software, supported by $104 million in total commitments.
Question: Can small AI models find zero-day vulnerabilities?
Yes, testing showed that small, open-weights models were able to recover much of the same vulnerability analysis as Anthropic's Mythos when tested against the same code samples.
Question: What is the 'jagged frontier' in AI cybersecurity?
It refers to the observation that AI capabilities in security do not scale smoothly with model size, meaning larger models do not always provide a proportional increase in discovery or analysis performance over smaller ones.
