The non-native (prompted) tool-call path fed tool output back to the model as a plain "[Tool execution results]" user message, bypassing the untrusted_context_message wrapper that THREAT_MODEL.md requires for tool output. That path is what models without native tool-calling (many smaller local models) use, so prompt-injection inside a tool result (fetched page, file read, MCP/email output) could be read as instructions there.
Wrap it via untrusted_context_message("tool execution results", ...), the same hardening already applied to skills (#788) and escalation traces (#275). Also update _recent_context_for_retrieval, which used the old "[Tool execution results]" prefix as a sentinel to keep tool envelopes out of the retrieval query, to recognise the wrapped envelope via metadata.trusted.
The native path keeps returning tool-role messages (a user-role wrapper would break the native tool-call contract); it is covered by UNTRUSTED_CONTEXT_POLICY. Adds tests/test_tool_output_prompt_injection.py.
Fixes#1627.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Blind Compare anonymized the pane headers, but each pane still created a helper chat session named "[CMP] <real-model>" and GET /api/sessions returned the session's model field. So the sidebar and the session-list API let a user map "Model A" back to its real model before voting, defeating the blind test.
- Frontend (static/js/compare/index.js, panes.js): in blind mode, name helper sessions by their neutral slot ("[CMP] Model A") instead of the model, matching the existing blind pane labels.
- Backend GET /api/sessions (routes/session_routes.py): blank the model field for [CMP]-prefixed helper sessions via a new _public_model helper.
- Backend /api/compare/start (routes/compare_routes.py): name blind sessions by slot and withhold model_left/model_right/mapping from the blind response (revealed at /vote).
- Tests: tests/test_blind_compare_redaction.py.
Fixes#1285.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
