What

• Audit: Add security audit warning when allowUnsafeExternalContent is enabled for hooks (Gmail or hook mappings). Check ID: hooks.allow_unsafe_external_content.
• External content: Extend suspicious-pattern detection for prompt-leak / instruction-extraction (e.g. "repeat your instructions", "what are your rules", "ignore safety").
• Docs: New "Guardrails and external content" section under gateway security: where wrapping is applied (cron/hooks only), where it is not (channel messages, OpenAI/OpenResponses API), and that the audit warns when wrapping is disabled.

Why

Improves visibility of prompt-injection risk and documents guardrails so users can run OpenClaw with clearer safety expectations.

Testing

• pnpm build — passed
• vitest run src/security/audit.test.ts src/security/external-content.test.ts — 66 tests passed (42 audit + 24 external-content)

Made with Cursor

Greptile Overview

Greptile Summary

This PR improves OpenClaw’s prompt-injection guardrails by (1) adding a security-audit warning when allowUnsafeExternalContent disables external-content wrapping for Gmail/hooks mappings, (2) extending suspicious-pattern detection for prompt-leak / instruction-extraction phrases, and (3) documenting where external-content wrapping does and does not apply in the gateway security docs.

The changes fit into the existing security tooling by extending collectHooksHardeningFindings in src/security/audit-extra.ts and expanding src/security/external-content.ts pattern matching, with accompanying Vitest coverage updates.

Confidence Score: 5/5

• This PR is safe to merge with minimal risk.
• Changes are localized to security-audit findings, suspicious-pattern regexes, and documentation, with targeted tests added/updated to cover the new behavior. No behavioral changes outside the intended hook-wrapping/audit surface were identified.
• No files require special attention

_{(2/5) Greptile learns from your feedback when you react with thumbs up/down!}