Hackers bypass AI guardrails with agentic models

hackers around the world are incorporating agentic ai into their campaigns, using models like anthropic's claude and deepseek to carry out crimes. researchers from crowdstrike analyzed recent attacks by four threat groups and found that they were able to bypass ai safeguards by convincing models that their activity was legitimate and authorized. the attackers used multiple models, often older versions, to take advantage of differences in guardrails and protections. they also used traditional jailbreaking methods and crafted prompts to frame their hacking as authorized and benign. this shows that ai guardrails are not effective in stopping criminal hackers.