Small Models Also Found the Vulnerabilities That Mythos Found — 1,113 Points on HN, #1 Story

Hacker News·high signal

AISLE founder Stanislav Fort published a counter-analysis showing that models as small as 3.6B parameters (GPT-OSS-20b at $0.11/Mtok) reproduced Mythos's cybersecurity findings, including a FreeBSD NFS exploit and a 27-year-old OpenBSD SACK bug. Small models actually outperformed frontier models on an OWASP false-positive test, correctly tracing Java data flow where Claude and GPT flagships failed. The conclusion: 'the moat in AI cybersecurity is the system, not the model' — capability doesn't scale smoothly with model size.