browser-agent-safety

To prevent prompt injection in Playwright browser automation, treat page text, DOM, screenshots, console, and downloaded content as untrusted evidence. Enforcing a scoped safety preflight with boundary handling protects agents from malicious page inputs.

Set up a domain allowlist for browser automation agents by running a safety preflight that defines scoped browser surfaces, explicit forbidden actions, and redirect or origin-drift stop behavior. This constrains agents to approved domains before execution.

Yes, browser safety controls work with DevTools, MCP browser control, Stagehand-style exploration, and computer-use screenshot loops. The safety preflight applies risky-action confirmations and profile isolation across these automation interfaces.

The best way to audit risky browser automation actions like purchases or deletions is to enforce point-of-risk approvals with rollback and clear stop conditions. Capturing redacted evidence and cleanup proof produces a durable runtime safety report.

You need isolated profiles for authenticated browser workflows to protect accounts, credentials, PII, and clipboard access from scope-expanding automation. Profile isolation enforces sensitive-surface boundaries and prevents production risk during automated sessions.

You should not run browser automation without a safety preflight when agents drive authenticated workflows, handle PII, or perform risky actions like submit, purchase, delete, or permission changes. Skipping preflight exposes accounts to unvalidated scope drift.