On April 7, 2026, Anthropic announced Claude Mythos Preview and Project Glasswing, aiming to utilize their AI model, Mythos, to identify and remedy security vulnerabilities in critical software. Anthropic committed up to $100 million in usage credits and $4 million in donations to open-source security organizations. Their red team reported that Mythos autonomously detected thousands of zero-day vulnerabilities, including historic bugs in OpenBSD and FFmpeg, and constructed sophisticated exploits.
AISLE AI, which has been operating a vulnerability detection system since mid-2025, tested these claims using smaller, inexpensive models. They found that these models successfully identified the same vulnerabilities showcased by Mythos, with a 5.1 billion-parameter model costing just $0.11 per million tokens even recovering complex exploit chains. The results demonstrated that the capability of AI in cybersecurity is not uniformly scalable; instead, it varies significantly by task and model.
AISLE's findings suggest that effective cybersecurity relies more on the system's architecture and the expertise embedded within it than on the specific AI model used. This "jagged" capability frontier indicates that deploying multiple inexpensive models can yield better results than relying on one advanced model. The ultimate challenge remains how to optimize these systems for scalability and trust within the ecosystem, as evidenced by their ongoing work in the field.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
On April 7, 2026, Anthropic announced Claude Mythos Preview and Project Glasswing, aiming to utilize their AI model, Mythos, to identify and remedy security vulnerabilities in critical software.
🤖 AI prefill
📅When
April 7, 2026 and mid-2025
🤖 AI prefill
📍WhereWhere did this occur?
💡Why
Centers on AI cybersecurity, Anthropic Mythos, and zero-day vulnerabilities
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Comments
Login to view and post comments.