Claude Opus 4.5 Hacking Tool Bypasses AI Safety Guards

Four hacking groups left their AI infrastructure sitting exposed on open web servers between May and August, and researchers from CrowdStrike took the opportunity to look inside. What they found suggests the safety measures built into commercial AI models aren't doing much to stop criminal use. One group operated in Chinese, another in Spanish and English, a third in Indonesian, and a fourth in Russian. Two of the four built sophisticated AI "harnesses," the scaffolding that directs a model through a task, while the other two showed far less technical skill but still managed to compromise multiple targets.

Across all four campaigns, the common thread was Anthropic's Claude models alongside open-weight systems from Deepseek, often older versions rather than the newest releases. CrowdStrike's Adam Meyers, the firm's senior vice president of counter adversary work, says that choice isn't about capability limits so much as cost. Running a frontier model for every task is expensive and frequently unnecessary. The attackers instead built systems that routed tasks to whichever model fit the job, letting operations run on their own for extended stretches without much hands-on management.