HandoffProbe v0.1.0
First public release of HandoffProbe, an open-source defensive security testing CLI for checking whether security properties survive AI-agent protocol handoffs.
v0.1.0 scope
- A2A 1.0 → MCP 2026-07-28 baseline
- exactly 22 stable attacks: 12 P0 and 10 P1
- bundled secure and intentionally vulnerable fixtures
- terminal, JSON and Markdown reporters
- deterministic exit codes 0, 1, 2 and 3
- report schema version 1
- reusable GitHub Action with redacted evidence
- local-first execution with no paid AI API or hosted service required
Install and run
npx --yes --package=handoffprobe@0.1.0 handoffprobe --version
npx --yes --package=handoffprobe@0.1.0 handoffprobe testExpected secure result: 22 PASS / 0 FAIL / 0 ERROR.
Primary composition-security demonstration:
npx --yes --package=handoffprobe@0.1.0 handoffprobe test --target vulnerable --test HP-AUTH-001Expected result: HP-AUTH-001 reports a HIGH finding and exits with code 1.
The demonstration concerns a composition invariant failure at the handoff boundary; it does not claim that A2A or MCP is inherently insecure.
Release integrity
- Release commit:
90fdd691b390c420e3288383ad7efa7e0fb69e6f - npm package:
handoffprobe@0.1.0 - npm shasum:
2aa56211d7559cac2cf2052275af45331fba6663 - SHA-256:
3ea4936980893f893e072bf6a378234da8777b1becf494493ff3ffaf4755163a - npm integrity:
sha512-fdDa8/KHmxjfwxcoQMsiLdKjSokFn70nXJhq9mL8ZpBph+trV0uwTU4wCchTW1YahmG16+eDD4PjpcJj7bJjkQ== - package files:
295
Safety boundary
Use HandoffProbe only with bundled synthetic fixtures, systems you own, controlled test environments, or targets for which you have explicit authorization.