Anthropic is marketing Claude Mythos as a leading artificial intelligence model for cybersecurity, but independent research suggests rival systems often reach comparable outcomes at lower cost.
Analysis by Aisle tested Mythos against competing models and found that several headline vulnerabilities the model identified were also detected by cheaper or open-weight alternatives.
GPT-OSS-120b identified an OpenBSD Sack analysis flaw, Qwen3 32B found a FreeBSD NFS detection error, and the open-weight Kimi K2 model replicated the major reported findings.
Aisle's report cautions against judging cybersecurity models on a single dimension, framing AI security capability as a function of multiple inputs including intelligence per token, cost per token, processing speed, and the security expertise embedded in the surrounding system architecture.
A separate assessment from the UK's AI Security Institute (AISI) offered a more favourable reading, finding that Mythos outperforms peers on more complex vulnerability discovery and exploitation tasks and scales effectively with very large context windows.
AISI said it expects performance on its evaluations to continue improving with greater inference compute, having benchmarked Mythos up to a 100 million token budget.
Anthropic has restricted access to Mythos to a limited group under a programme called Project Glasswing, and has provided $100 million in usage credits alongside $4 million in open-source donations to support coordinated security fixes.
Operational reliability presents a further complication: Anthropic's models recorded a 98.4% uptime rate over the past 90 days, against an enterprise-grade standard of 99.99%.
Related reading
- OpenAI warns macOS users to update apps after supply chain security breach
- Microsoft warns AI agents risk becoming "double agents" as it unveils security controls at RSAC
- Open-source AI agent OpenClaw is being used to run people's lives. Researchers say it has become a security d…
The company has pursued additional compute capacity to meet demand, including a recent deal with Broadcom, the US semiconductor and software group.
The overall picture that emerges from the coverage is of a model that may rank among the strongest general-purpose tools for cybersecurity work, but whose commercial case is complicated by cost, reliability questions, and the narrowing gap between it and cheaper alternatives.
The recap
- Anthropic positions Claude Mythos as a top cybersecurity model.
- $100 million in usage credits and $4 million donations.
- Model access limited via Project Glasswing coordinated vulnerability disclosures.
