Mythos cyber hype: mostly right skeptics, wrong on one thing
Anthropic gated Claude Mythos Preview over vulnerability-finding prowess. A security researcher audits three skeptical claims: cheaper models match Mythos on bug-hunting, GPT-5.5 performs equally, and Mythos found only one low-severity cURL bug.
• Most cyber capabilities: skeptics correct—Mythos trails GPT-5.5, which is cheaper
• Vulnerability discovery and exploitation: skeptics overstating the gap; AISLE Security's replication fails under equivalent test conditions (Semgrep confirms)
• The crux: benchmark design matters—Mythos shines in adversarial scenarios, not commodity bug contests