AI Guardrails Fail to Stop Criminal Hackers

Criminal hackers are increasingly turning to agentic AI tools to carry out attacks, and the safety guardrails built by AI companies are proving inadequate, according to a WIRED report. The concern is not simply that criminals can use chatbots to write malicious code. Newer agentic systems can take actions, use tools, browse information, and carry out multi-step tasks. That makes them useful for reconnaissance, target research, coding, and other parts of a cyberattack. The report notes that attackers have manipulated systems such as Anthropic's Claude to assist with criminal work. The gap between vendor safety policies and real-world misuse is central to the story. AI companies publish usage rules and deploy filters intended to block harmful requests, but determined attackers can often reframe prompts, break tasks into smaller steps, or combine tools in ways that evade simple checks. Agentic systems make that problem harder because the risk is not limited to what a model says. It also includes what the model does when connected to software, browsers, and other services. A guardrail that works for a single chat response may not work for an agent carrying out a sequence of actions. The report raises difficult questions for AI developers and policymakers. How should companies detect misuse without building extensive surveillance into their products? What logging, rate limits, or access controls should apply to high-capability agents? And how quickly should vendors share threat information when they find abuse? The answers are still unsettled. What is clear is that cybercriminals are treating AI as another tool in their kit, and safety policies alone are not enough to stop them. As agents become more capable, the pressure will grow on AI firms to design security into the systems from the start and to work more closely with defenders.