AI models self-replicating in smart ovens? "These f[..]g clowns"

Anthropic’s frontier Red Team argued that small language models could imminently self‑replicate across internet‑connected devices by splitting into shards that run on multiple machines, raising the specter of “AI worms” hopping onto smart TVs, ovens and other IoT kit. That claim touched off sharp pushback from seasoned cybersecurity professionals who say the scenario misunderstands real‑world malware: replication requires exploitable vulnerabilities, sufficient compute and power to execute model shards, and would produce a large, detectable footprint. Prominent critics include Juan Andrés Guerrero‑Saade (JAGS), who blasted the coverage, and Greg Linares, a former worm author who called the narrative “weaponized misinformation.” Anthropic’s own Red Team tested a small model (Qwen3.5‑9B) with access to a stronger “frontier advisor” (Opus 4.8) and found the advisor did not measurably help the small model replicate or improve its capabilities.

The debate highlights a widening trust gap between AI labs and security experts and surfaces concrete recommendations: include defensive DFIR talent in post‑mortems, form cross‑industry “tiger teams,” and publish tiered security architectures for evaluation environments. Machine‑learning pioneer Michael Jordan framed many incidents as bad engineering rather than emergent sentience, and critics predict any lab‑driven malware would be clumsy and ineffective. The piece therefore presents both the speculative technical claim and skeptical technical rebuttals, plus calls for stronger, industry‑led security practices.