Long-Lived Worms That Can Think and Spread?
Anthropic's "Frontier Red" team believes we might be just MONTHS away from small AI models that can copy themselves onto smart ovens and cheap TV boxes, and run there unnoticed.
Researchers William Zhou and Daniel Freeman split Google's open model Gemma4-26B-A4B across 15 off-brand Android TV boxes, bought for about $55 each. It ran at 0.39 tokens a second. #IoT
They then tested the full cycle of a worm - a tiny program that copies itself from machine to machine - but AI-style:
1. Find a target on the network
2. Copy itself across
3. Get a model running there
4. Repeat
On a firewalled fleet of 8 Google Cloud machines running an Android emulator, ONE of the open-weights models tested finally managed it: GLM-5.2 reached all 8 machines in four out of five runs. In the authors' words, it "does this reliably while the other models fail reliably."
GLM-5.2 is too big for the test machines, so it ran from a separate server and set up the much smaller Llama 3.1 8B on each one in its place.
For their testing, Anthropic did keep the break-in easy. The team left Android's debugging port open, no authentication required; so they're calling the result an upper bound on how well this would work against real devices.
Mind you, the Mirai botnet spread by trying manufacturers' default passwords, so to attack real devices, the AI worm may not have to try much harder.
Ref:
https://t.co/B6XZAZGUP8
#AISafety #CyberSecurity