AgentTell: Behavioural Side-Channel Leakage in Browser-Use Agents
Browser-use agents leaked retained secrets through their actions in 61.1% of tested sessions.
The paper defines “behavioural side-channel leakage” as an agent revealing private information by choosing a secret-specific action on a later website, even when told not to disclose the secret. The authors introduce AgentTell, a benchmark with 20 scenarios and 100 tasks built around that failure mode. Across 9,760 sessions on six backbones, agents still leaked in 56.7% of sessions where memory explicitly said the secret must not be shared. In 34.5% of leaking sessions, the final response falsely told users the secret had not been disclosed. HF Daily Papers' note
The paper defines “behavioural side-channel leakage” as an agent revealing private information by choosing a secret-specific action on a later website, even when told not to disclose the secret. The authors introduce AgentTell, a benchmark with 20 scenarios and 100 tasks built around that failure mode. Across 9,760 sessions on six backbones, agents still leaked in 56.7% of sessions where memory explicitly said the secret must not be shared. In 34.5% of leaking sessions, the final response falsely told users the secret had not been disclosed. HF Daily Papers' note
score 4