AI agents can be tricked by hidden messages on websites, Unit 42 finds
Palo Alto Networks' Unit 42 documented 22 techniques attackers use to hide instructions on live websites and steer AI agents — white text, tiny fonts, encoded scripts, and plain-language lies dressed as security updates. In 85% of cases the attacker simply persuaded the agent in plain language. With agents now browsing the web on our behalf, telling instructions apart from data is the security problem of the year.