Repository navigation
Description
Is your feature request related to a specific problem?
The first-party ModelArmorPlugin documents that tool output is not screened:
Tool results arrive as user content whose only part is a function_response, which the plugin does not send to Model Armor. Agentic pipelines that ingest external text via tools (issues, email, web search, MCP) often need screening on tool output, not just user prompts and model replies.
Describe the Solution You'd Like
Add a minimal sample under contributing/samples/integrations/model_armor_tool_output/ demonstrating a companion plugin on after_tool_callback that stringifies tool results and calls Model Armor's SanitizeUserPrompt API, registered alongside ModelArmorPlugin on the same App.
Impact on your work
We run a multi-tool agent control plane where the primary injection vector is hostile content returned as tool output. We use a custom after_tool_callback plugin today; a first-party sample would reduce unsafe copy-paste from older adk-samples safety-plugins code.
No critical deadline.
Willingness to contribute
Yes
Describe Alternatives You've Considered
- Using
ModelArmorSafetyFilterPluginfrom adk-samples: older callback shapes; not the ADK 2.8 first-party integration path. - Waiting for core plugin extension: larger API design (tracked separately as a follow-up feature request).
Proposed API / Implementation
class ToolOutputModelArmorPlugin(BasePlugin):
async def after_tool_callback(
self, *, tool, tool_args, tool_context, result, **kwargs
):
text = stringify(result)
if not text:
return None
# Call modelarmor SanitizeUserPrompt on text; return error dict if MATCH_FOUNDSample registers [ModelArmorPlugin(...), ToolOutputModelArmorPlugin(...)] on one App.
Additional context
- The limitation is documented in the adk-python integration guide under Limitations.
- I will open a PR referencing this issue the same day if maintainers are open to the approach.
Activity
sanketpatil06 commentedon Sep 1, 2026
Hi @saurabh4269, Thanks for bringing this and taking time to raise a PR. Your request is currently being reviewed, and we will let you know if anything further is needed. We appreciate your contribution.
wuliang229 commentedon Oct 2, 2026
We will enhance the ModelArmorPlugin instead of creating something new. Closing.
GWeale commentedon Oct 7, 2026
thanks @saurabh4269. we added this to ModelArmorPlugin instead of a sample: fdda17c (#6969) adds tool output screening, set up with tool_output_template_name. it's on main, not in a release yet. please open a new issue if this is still happening.