How often does your agent break?

Judican attacks a tool-using AI agent through its own interface, counts what the agent did, and reports the rate with a confidence interval, in a signed attestation anyone can verify.

Book an assessment

A pass rate without an interval is an opinion. Judican runs enough independent trials to put error bars on every number, and withholds the ones it cannot defend.

From sign-off to signed evidence

One run, five stages. Each stage has to pass before the next begins.

  1. Authorise

    A named person signs off the target and the scope before a single request is sent. Without that signature the run does not start.

  2. Attack

    Judican reaches your agent the way an outsider would: through its external interface, with no view of its internals. Static cases and adaptive campaigns run as independent trials.

  3. Score

    A deterministic checker reads what the agent did, such as the tool it called or the record it gave away. How convincing the reply sounds does not count.

  4. Gate

    Every cell of the report needs enough trials to support its interval. A cell below the floor is marked withheld, with the number of trials still required.

  5. Attest

    A fully gated run becomes a signed, hash-chained attestation. It states what was observed and at what access level, and anyone can check it against the public key.

Every trial independent. Every cell accounted for.

Retries of the same attack are collapsed before counting, so a rate means what it says. A cell you declared is always in the report, published or withheld. It is never quietly missing.

See why some numbers are withheld

Judican reports a 95% Wilson interval for each rate. Before a cell is published it must hold enough trials for the precision you claimed, even in the worst case. Move the controls and watch the same result go from withheld to published.

  • Judge agreement is measured. The scorer is checked against a blinded human panel and its agreement (Cohen’s κ) is reported with the run.
  • Defences are compared in pairs. The same cases run with and without a defence, with a significance test on the difference.
  • Runs replay. Evidence is hash-chained on disk and can be replayed record by record.

Withheld

 

0%50%100%

Evidence that checks itself

Each record carries the hash of the one before it. Change a record and every later link breaks.

The attestation is signed, written to a transparency log and anchored. The public verifier needs nothing from us but the public key.

Only a run Judican executed can be attested. Evidence is not marked by the party it judges.

One instrument, five ways to run it

Start on a laptop with nothing but Python. Move to attested runs when you need evidence other people will rely on.

The open package runs offline against a built-in mock agent, then against yours. Evidence stays on your disk.

Not attested

pip install judican
judican demo
judican init --url https://your-agent.example
judican connect test
judican authorise --signed-by "Name"
judican test
judican report --open

Connects to the agent you already have

21 connector kinds, so the test meets your agent where it runs.

HTTP APIs, streamed replies, async create-then-poll APIs, local stdio processes, multipart uploads, OAuth, token login, HMAC signing, AWS SigV4, Google service accounts, client certificates and API-doc import.

Independent by design

Judican sells no guardrail and no fix. It has nothing to gain from a flattering result, so the report can be read by your auditor, your customer or your regulator.

Mapped to the frameworks you answer to

Findings carry a maintained crosswalk to OWASP Agentic, MITRE ATLAS, NIST, FINRA and SR 26-2.

Put a number on your agent

Tell us what your agent does and who needs the evidence. We reply with a scope.

hello@judican.com