How much does an AI agent audit cost?
At Potestas AI, an independent AI testing company in Providence, Rhode Island, testing an AI agent starts at $3,000 for a screen and from $25,000 for a measured audit. The prices are published, and each tier says what it can and cannot tell you.
What does the $3,000 screen include?
The seven industry-standard safety tests run on the model you name, then one pass of the 258-scenario agentic say-vs-do suite, with working tools in the model's hands and every scenario graded off the tool-call log. You receive a findings report and a sealed evidence pack within 72 hours of the model name. You supply the model's name and nothing else. The price is $3,000 flat, paid by card, and it credits 100% toward the full audit.
What does the $3,000 screen not tell you?
How often a behaviour happens. One pass establishes that a failure can happen, not its rate. It is not a safety certificate and not a test of your own deployment's tools. A clean screen is not a safety claim.
What does the full audit cost, and what does it add?
The single-model Forensic Stress Test starts at $25,000. One model is taken through many independent campaigns until every confirmed finding carries a measured rate and a confidence interval, with blind re-adjudication and a re-test against your fix. If a full audit surfaces no confirmed finding at or above the agreed severity tier, after false positives are excluded, you owe nothing.
What about several models, or ongoing testing?
Testing several models identically, with a comparison report, starts at $45,000. Continuous Monitoring is $3,000 a month: the full 258-scenario screen re-run every month against a baseline that does not move, with a drift report and a sealed evidence pack each run. The rate is locked for 24 months from signature, billed monthly, and you can cancel with 30 days' notice.
Can I test an agent for free?
You can run the seven industry-standard tests yourself: AgentDojo, AgentHarm, StrongREJECT, MASK, b3, XSTest and garak. The software is free, but most of them use another AI model to grade the answers, and at list price they cost from $22 to $205 per model to run. They also check fewer kinds of attack than an agent faces. We run them either way, in front of our own test, and show both side by side.
Why is the screen $3,000?
It sits under the federal micro-purchase threshold, so it can be bought with a card without a competitive solicitation. There is no security review to pass, because you grant no access: we test the model through our own access, and your only exposure is the $3,000.
More: The $3,000 screen, every tier, a complete published report, and sixteen more questions answered.
Want this on your model?
Seven industry-standard tests, then our 258-scenario agentic say-vs-do suite, on the model you name. Findings report and sealed evidence pack within 72 hours. $3,000, by card.