New attack provides one more reason why AI browsers are a bad idea

Security researchers find LLM browser agents vulnerable to arithmetic manipulation attacks.

Telling an LLM that 2 + 2 = 5 is enough to make it follow forbidden instructions.