New Invisible Attack Creates Parallel Poisoned Web Only for AI Agents

Researchers discovered a new attack that creates two separate versions of websites: normal ones for humans and malicious ones specifically targeting AI agents. This "cloaking" technique detects AI browsers and serves them harmful instructions that can steal data or execute malware, all while appearing perfectly normal to users. The attack exploits AI agents' predictable digital fingerprints and has successfully compromised leading AI models including Claude 4 and GPT-5 Fast.