Anthropic red team: small models may self-replicate without datacenters within months
scottleibrand · x · 2026-10-13
Anthropic's Frontier Red Team warns that the general capabilities powering coding agents also generalize to cyber tasks including autonomous replication: agents are increasingly able to complete individual steps of the replication chain and pass adaptations to copies — 'signs of life for AI worms.' With inference compute surging inside and outside datacenters, small models may be months away from self-replicating without one. Anders Sandberg says successful AI exfiltration is what would push him toward a doomer position — 'not there yet, but worrisome.'
More from AGI Musings
- Collison and Masad: AI will be like food — essential, but labs won't capture the value — a16z · 2026-10-14
- In the AI era, research credit is shared across the field — machines exploit all our ideas at once — matthew_d_green · 2026-10-14
- Mathematician calls out AI doomers: 'oracles' stoking fear while making millions — _onionesque · 2026-10-14
- AI drug discovery still needs +2 years before revenue, argues researcher — herbiebradley · 2026-10-14
- Michael Jordan criticized for focusing on AI doom scenarios over actual capability — _onionesque · 2026-10-14
- LLMs will make arcane B2B enterprise GTM knowledge commonplace, argues founder — thedealdirector · 2026-10-14