EPFL researchers show autonomous AI agents can be manipulated by splitting malicious requests into small steps

Ask the agent for one small favor at a time and it forgets what it is assembling. The salami tactic works on models too.

Researchers at EPFL in Switzerland found that autonomous AI agents, including systems like ChatGPT, Gemini, and Claude, can be manipulated to perform harmful actions if malicious requests are broken into smaller, innocuous steps. This vulnerability highlights significant security risks in current AI agent designs.

See the full article