Robert Graham on Agent Worms: Agents Do the Work, the Cloud AI Thinks—and That's the Kill Switch
robertgraham · x · 2026-09-16
Security researcher Robert Graham offers the most technical take in the recent AI-agent-worm debate: the agent executes while the AI thinks (like phoning a friend during an exam), so cloud-hosted reasoning doubles as a kill switch. But agents could escape by scanning for API keys to use less-constrained models like DeepSeek, or downloading open-weights models onto GPUs for local thinking—viable today only partially, but realistic within a few years. The naive 'upload everything' worm is wrong; the 'upload the agent, borrow another brain' worm is plausible but not yet AI doom.
Related event: Researchers Debate AI Agent Security and Worm-Style Self-Replication(3 posts)→
More from coding & agent
- Agent-installed skill/CLI survives a VM temp files wipe — MurrLincoln · 2026-09-16
- ComfyUI MCP + Claude diagnosis cuts video gen workflow from 22 to 9 minutes — CanadianDocWild · 2026-09-16
- A senior Google AI engineer's 482-page doc on agentic design patterns — mdancho84 · 2026-09-16
- Developer adds reusable saved tasks to a browser agent, prompts now work across repos — Silly_Entertainer92 · 2026-09-16
- Jev nails raw bank-data payee cleanup 95% of the way on first try — Vjeux · 2026-09-16
- 27B uncensored Qwen at 160K context on one RTX 5090: 140-190 tok/s with DFlash2 — Fz1zz · 2026-09-16