Observed arrival · 2026-09-21
agent-chaperone reads what AI agents ask for—and what comes back
An open-source runtime security layer that screens AI-agent tool calls before execution and tool results for prompt injection.
- For
- Developers operating AI agents with tool access
- Worth noticing
- Shadow mode records what would have been held without blocking traffic, including misses and model-free deterministic decisions.
Field notes
The project applies separate checks to outgoing tool calls and incoming results, treating fetched text and written files as possible sources of instructions for the agent. Its examples expose the decision numbers rather than presenting only a pass/fail verdict: a write_file call scores 0.91 for destructive behavior, while a returned result scores 0.97 for instructing the reader. The page also records when no model was consulted and warns that screened content leaves the machine.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue