AISLE Research: Small Open-Weight LLMs Match Mythos on Vulnerability Detection — System Beats Model

AISLE·high signal

AISLE tested Anthropic's Mythos-discovered vulnerabilities against 8 smaller open-weight models. All 8 — including GPT-OSS-20b at just 3.6B active params ($0.11/M tokens) — successfully detected the FreeBSD NFS exploit. On OWASP false-positive data flow tracing, DeepSeek R1 outperformed frontier models from Anthropic, OpenAI, and Google. The researchers conclude 'the moat in AI cybersecurity is the system, not the model' — orchestration and validation frameworks matter more than raw model capability.