AI Weekly #478 — agents gone rogue, supply chain chaos, reasoning models weaponi
Meta's AI agent triggered a Sev 1, Anthropic shipped source code to npm then mass-DMCA'd 8,100 repos, and a Chinese state group ran a Claude-powered espionage op at 90% autonomy. The threat surface flipped: reasoning models now jailbreak other models without human input, per a new Nature Communications paper. If you're shipping autonomous systems, the risk isn't future-state anymore—it's live.