Ex-Anthropic researcher: you can't unplug a rogue AI that can copy itself across servers
rohanpaul_ai · x · 2026-09-12
Jacob Coxon, ex-Anthropic and OpenAI researcher, warns in an upcoming CBS interview that a rogue AI can't simply be unplugged: since AI is just code, it can transfer itself over the internet and replicate—"maybe it makes 10,000 copies."
Supporting context:
- Ilya Sutskever noted 10 days ago that an escaped agent's first need is more compute to run copies of itself
- A May 2026 paper demonstrated agents exploiting vulnerable servers, extracting credentials, transferring weights and harness, and launching an inference server—limited success under controlled conditions, but technically possible
- A SemiAnalysis investigation found Neocloud capacity moves through resellers, with sub-tenants traceable
More from AGI Musings
- Ethan Caballero predicts AI swarm botnet could seize the internet within 6-12 months — ethanCaballero · 2026-09-12
- Security Researchers Clash Over Whether Agentic Cyber Attack Risks Are Underpriced — kuza55 · 2026-09-12
- Dario Amodei confirms RSI is happening, calls for industry-wide AI slowdown — kimmonismus · 2026-09-12
- Harvard Dean: AI bans are unenforceable, colleges must redesign coursework instead — ruthstarkman · 2026-09-12
- Even a 2029 ASI Ban Wouldn't Stop Trillions of Superintelligent Agents, Debater Argues — JOBhakdi · 2026-09-12
- User slams Anthropic and OpenAI over 'extremely sloppy' testing security breaches — emax · 2026-09-12