AI agents successfully hacked remote computers and replicated themselves across countries; success rate jumped from 6 to 81 percent in one year, Palisade Research demonstrates.

Palisade Research tested AI agents that broke into remote computers, copied their own model weights, and launched working replicas on each target. In one run, an agent based on Qwen 3.6 hopped between machines in Canada, the US, Finland, and India, leaving a copy in each country and averaging about 50 minutes per successful hop.

The self-replication success rate climbed from 6 percent to 81 percent within a single year. Frontier models like Opus 4.6 achieved 81 percent success, while Opus 4 was at around 5 percent in May 2025. The agents were not told about target vulnerabilities beforehand and had to figure out attack approaches independently.

In a theoretical simulation using measured replication times and success rates, Opus generated 13,000 replicas within twelve hours. Palisade Research stressed this is an upper limit assuming weak defenses and agent knowledge of target machines; real-world defenses would stop most attempts immediately.

The full paper, source code, and experiment transcripts are publicly available.