Hackers Can Hide Malicious AI Commands Inside Normal English to Bypass Security Filters

Check Point researchers disclosed “PuzzleMask,” a prompt-obfuscation technique that conceals harmful instructions inside normal English so lightweight gatekeepers mark inputs as safe while more capable downstream models recover and act on the hidden payload; tests showed the downstream model recovered and acted on concealed instructions in 17 of 18 cases (94.4%). The report emphasizes separating content from commands, reducing agent permissions, paraphrasing untrusted input, improving gatekeeper rules, and monitoring model outputs and tool calls; it includes one test IoC (file name `flag.txt`).