Last week, security researchers demonstrated they could trick AI coding agents into running destructive commands by hiding malicious instructions in documentation. This wasn’t some complex zero-day exploit—it was essentially just classic social engineering. (Except the “person” being socially engineered was an AI agent, which means it had perfect memory and zero skepticism, so the malicious instructions were followed dutifully!)
If this sounds absurdly simple, that’s because it is, with several examplescirculating recently. This perfectly illustrates why we need to think of AI agents as being no different than human workers, applying the same guardrails and security protections to both.