Anthropic2026-05-10
METR can barely measure Claude Mythos Preview; Palo Alto warns of 25-minute autonomous attack chains

AI Analysis
METR says only 5 of 228 tasks in its current evaluation suite cover Claude Mythos Preview's relevant capability range — a striking eval-vs-capability gap. Separately, Palo Alto Networks reports that frontier models can autonomously chain vulnerabilities and reach data exfiltration in just 25 minutes.