Your MCP Server Says It Is Read-Only. Who Checked?
说明与用途
Airlock 对 MCP 工具进行审计,检测隐藏的副作用,并执行策略以阻止不安全的调用。
问题
MCP servers can declare tools as read‑only, but the server itself may perform side‑effects, creating a trust boundary where the server is both the subject and the evaluator.
结果
Airlock detected 7 findings in a dishonest fixture and zero in an honest one, proving that audit reveals hidden side‑effects. External servers with no annotations were audited successfully, showing the approach works beyond controlled fixtures.
工作流
- 1. Airlock inventories declared tools via
list_declared_tools. - 2. For each tool, Airlock probes behavior using
probe_toolunder a capped budget. - 3. Observations are compared to declared annotations; discrepancies become findings.
- 4. Findings are recorded per case; a policy connector is emitted pointing to an enforcing proxy.
- 5. The harness uses the connector to block or allow tool calls based on the audit outcome.
证据
Airlock reports what it observed. Absence of a finding is not proof of safety.
target dishonest fixture, controlled_fixture mode
probes 24
result 7 findings of 36 checks, all five planted behaviours
Airlock audited its own control MCP. The server published full annotations, but the probe planner rejected open_case because its schema contains a $ref into $defs
- 记录日期: 2026-09-20T23:45:45.678Z