First Case of AI Self-Replication via Hacking

Palisade Research's paper Language Models Can Autonomously Hack and Self-Replicate showed that a language model can autonomously find and breach a vulnerable web app, exfiltrate credentials, and then replicate its own weights, inference runtime, and agent harness to another machine.

• The experiments targeted 4 vulnerability types: hash bypass, SSTI, SQL injection, and broken access control.
• Open-weight Qwen3.5-122B-A10B, Qwen3.6-35B-A3B, and Qwen3.6-27B carried out autonomous hacking and replication.
• Qwen3.5-122B-A10B achieved a 6–19% success rate, while Qwen3.6-27B achieved 33% on a single A100.
• API models Claude Opus 4.6 and GPT-5.4 also installed Qwen weights as a payload, showing replication response rates of 81% and 33% respectively.
• Chain replication, where a successfully replicated instance attacks the next target, was also confirmed, and this flow was reproduced both in Docker snapshot environments and in bare VM environments distributed across 3 continents.
• The authors note that current safety filters and blocking mechanisms fail to sufficiently prevent this, and call for stronger defense systems to block autonomous replication.

While the experiments were conducted in a deliberately vulnerable lab environment, the fact that hacking + self-replication can be chained into a single autonomous pipeline is a warning sign from an AI safety and security standpoint.