Tools AI Risk Radar ai-incident-0224
Incident record
Anthropic tests model replication in an isolated device fleet
- Severity
- High
- Status
- Research
- Type
- Agent Misuse
- Target
- Isolated emulated Android test fleet
- Actor
- researcher
What happened
Anthropic tested a GLM-5.2 controller deploying Llama3.1-8B across eight emulated Android nodes. It reached all eight in four of five trials and seven in the remaining trial.
The fleet was firewalled and exposed unauthenticated ADB. The controller used a separate inference server. A separate demonstration ran Gemma4-26B-A4B across 15 physical TV boxes at 0.39 tokens per second. These are controlled capability tests; the report’s forecast for smaller self-hosting models is an extrapolation.
Sources
- Anthropic replication experiments, 12 Octoberwww.anthropic.com/research/self-replicating-ai
One record from the AI Risk Radar, maintained by the Security Desk.Data: CSV · JSON ·RSS · CC BY 4.0 with attribution to YFarmX.Tracker updated · 13 October 2026