Problem. Snapshots and screenshots of signed-in pages are stored in session events, and archives are written to agent memory. Both are readable at "reader" level, so on a public agent anyone can read them. The docs claim they are private.

Fix. Refuse browser access on public agents. Redact browser tool inputs and outputs for everyone except the owner, including if the agent becomes public later. Keep archive copies out of reader-visible memory.

Branch: feat/integrated-browser