Root cause (from infra-lead PR#7 review id=724):
Sanitization in PR#7 wrapped peer text in [A2A_RESULT_FROM_PEER]
markers, but the markers themselves were not escaped — a malicious
peer could inject "[/A2A_RESULT_FROM_PEER]" to close the trust
boundary early, making subsequent text appear inside the trusted zone.
Fix:
- Create workspace/_sanitize_a2a.py (leaf module, no circular import
risk) with shared sanitize_a2a_result() + _escape_boundary_markers()
- _escape_boundary_markers() escapes boundary open/close markers in the
raw peer text before wrapping (primary security control)
- Defense-in-depth: also escapes SYSTEM/OVERRIDE/INSTRUCTIONS/IGNORE
ALL/YOU ARE NOW patterns (secondary, per PR#7 design intent)
- Update a2a_tools_delegation.py: import from _sanitize_a2a; wrap
tool_delegate_task return and tool_check_task_status response_preview
- Add 15 tests covering boundary escape, injection patterns, integration
shapes (workspace/tests/test_a2a_sanitization.py)
Follow-up (non-blocking, noted in PR#7 infra-lead review):
- Deduplicate if a2a_tools.py also wraps (currently handled in
delegation module only — callers get sanitized output regardless)
- tool_check_task_status: consider sanitizing 'summary' field too
Closes: molecule-ai/molecule-ai-workspace-runtime#7 (wrong-repo PR
that this supersedes)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
