Lab
In auditAgent assurance harness
A research thread on testing whether an agent's effective authority stays inside its approved authority.
Not yet audited
- Capabilities
- Agent evaluation · Security testing · Assurance
- Evidence
- confirmed
- Disclosure
- general
Status
Not yet audited. The harness is a working idea, not a published tool. It is listed here so the direction is visible without pretending that an unaudited prototype is evidence.
Working question
An agent can be given a narrow tool and still compose several safe calls into an unsafe outcome. The question behind this thread is whether an assurance harness can test the authority an agent actually exercises, rather than the authority described in its prompt or policy.
That means exercising the workflow, not only the model: which tools were called, in what order, with whose credentials, and what happened when a call failed. The output should change a deployment condition or a review threshold. Otherwise the harness is another report.
Boundaries
No client systems, no production credentials and no claims of coverage. The next step is a repository audit and a small, reproducible test set before any public build is linked.