AgentShield is a security tool — it exists to stop sensitive-data leakage through LLM agent tool calls. Findings about its own defenses are taken seriously.
Only the latest release line (v1.6.x) is supported. Fixes land on main
and ship as patch releases; see CHANGELOG.md.
Use GitHub's private vulnerability reporting:
Repository Security tab → Report a vulnerability (direct link: https://github.com/liuhaolin07/AgentShield/security/advisories/new)
Please do not open a public issue for a vulnerability.
Include:
- the version / commit you tested,
- a minimal reproduction (tool call, payload, policy),
- the expected vs. observed decision (blocked vs. allowed).
In scope:
- a secret reaching a tool call without being flagged (scanner bypass),
- fail-open behavior under malformed input where fail-closed is promised,
- audit-log integrity issues (see the audit notes in the README).
Out of scope:
- false positives on benign content — report those as a normal issue,
- the simulated
send_httptool making no real network request (by design).
Best-effort response, typically within a week. This is a small project with no bug bounty; reporters are credited in release notes unless they prefer otherwise. Please allow time for a fix before any public write-up.