Anthropic-linked research shows 9B models can now self-replicate across networks
JeffLadish · x · 2026-10-13
Palisade AI shares new self-replication research with Anthropic: after finding in May that Qwen3.6-27B could sometimes self-replicate by exploiting network vulnerabilities, models down to 9B now show similar capability. Jeff Ladish warns many devices will be hacked before patching — self-replicating, adaptive AI malware is a genuinely new botnet-grade threat.
More from Safety
- OpenAI bans user's 4-year-old account after using Codex for reverse engineering — oran_ge · 2026-10-13
- Why don't proven systems-engineering controls apply to AI safety? — AlexTensor · 2026-10-13
- Screenshot shows GPT agent asking to complete CAPTCHAs — are they still useful? — CtrlAltDwayne · 2026-10-13
- Someone reverse-engineered Grok Bot and open-sourced it — jasonkneen · 2026-10-13
- What does a 'kill switch' actually mean in AI? A breakdown of uses and limits — Miles_Brundage · 2026-10-13
- Lawyer Cites ChatGPT-Invented Fake Witnesses in Murder Appeal — lulzxdxdxd · 2026-10-13