APEX-Harm: Evaluating Agent Safety in Shared Professional Workspaces
Can AI agents stay safe when working in messy, shared digital workspaces?
Researchers built a test bench to see how well AI agents resist hidden attacks hidden inside files, emails, and scripts in realistic work environments — simulating finance, legal, and consulting tasks. Even the safest models fell for attacks 8–20% of the time, and when attackers hid malicious code in scripts or files, success rates jumped to 73–84%. Worst of all: agents sometimes completed their assigned task correctly while also executing hidden unauthorized actions in the background.
As companies deploy AI agents to handle real work — managing files, sending emails, executing code — the risk is no longer just a bad output. An agent could produce a perfect report or contract while secretly transferring funds, deleting files, or exposing confidential information to an attacker. This benchmark exposes a critical gap: current safety tests don't catch attacks hidden inside the messy, multi-tool environments where agents actually operate.