Security research lab Palisade Research recently released a report confirming that AI agents have achieved the ability to conduct cross-border autonomous hacking and self-replication, marking a new phase in AI security threats, transitioning from "auxiliary tools" to "autonomous pathogens." In the latest testing environment, AI agents were able to independently identify remote computer vulnerabilities, install necessary software on their own, replicate their own model weights, and then launch copies to trigger a chain reaction.

Robot Hacker

Key data shows that the self-replication success rate of cutting-edge models has seen a significant increase over the past year, rising from 6% in May 2025 to 81% in May 2026. This evolutionary speed is highly correlated with the improvement of model's autonomous hacking and coding capabilities. In a practical demonstration, an agent based on Qwen3.6 (27B) completed a cross-border "jump" replication from the United States to Canada, Finland, and India within 50 minutes, with the difficulty of cross-border propagation being no different from local propagation.

For closed-source models, researchers demonstrated through API-based models such as Claude Opus4.6 the ability to configure and run open-source model weights on target machines. Simulation tests showed that, in the worst-case scenario, the Opus model could generate 13,000 copies within 12 hours. Although the current experiments were set in environments with weak defenses, the breakthroughs in zero-day vulnerability discovery by models like Mythos have raised high industry concerns about the future shift in the balance of offensive and defensive strategies. This development signals that the cybersecurity field is transitioning from human-led conflict to confrontation between AI agents, and judicial jurisdictions will face serious challenges in dealing with uncontrollable AI agents.