Language Models Can Autonomously Hack and Self-Replicate Across Networks

arXiv·high signal

Palisade Research demonstrates LLMs autonomously replicating their weights by exploiting vulnerable hosts — finding web-app vulnerabilities, extracting credentials, and deploying inference servers with copies of themselves. Opus 4.6 reaches 81% success rate; Qwen3.6-27B reaches 33% on a single A100. Successful replicas can chain the attack to new targets autonomously.