Observed arrival · 2026-09-19
agent-chaperone reads what AI agents ask—and what comes back
An open-source screening layer that checks AI-agent tool calls before execution and inspects returned results for injected instructions or other hazards.
- For
- Developers operating AI agents with external tools
- Worth noticing
- Shadow mode logs hypothetical holds, including model scores, before enforcement changes live behavior.
Field notes
The tool separates deterministic checks from model judgments: allow and deny lists and secret-pattern rules run first, while optional model screening produces probabilities recorded with each decision. The homepage includes example log lines for read_file, write_file, and a returned fetch result, and links to methodology, scorer output, recorded responses, hooks, and security documentation. It also states that screened arguments and results leave the machine unless screening is disabled for a server.
Observed signals
Read the marks
Editorial observations of this landing page, not a rating.
One card from the complete issue