AI Agent Security Review
AI Transparency & Human Oversight Review
Check that AI disclosures, approvals and human handoffs work in the deployed workflow.
Who it is for
For patient communication platforms, telehealth services and businesses that offer AI interactions or AI-assisted decisions. It is a defined engagement within our AI Agent Security Review.
Best fit: before you launch or expand an AI interaction, after you change an approval step, or when a customer or policy asks you to show that disclosure and human review work as described.
The review tests what people actually see and what the system actually does, not what the design document says.
What the review can examine
We select the modules that apply to your system and agree them in writing. They can include:
- Disclosure behavior across web, mobile, chat, email or voice.
- Whether the human-contact routes you promise actually reach a person.
- Whether a required approval happens before content is released or an action runs.
- Whether that approval covers the exact content or action that is executed, not an earlier version.
- Alternate paths that bypass approvals or other controls.
- Intervention, revocation and safe-stop behavior, where relevant.
- Marking of generated content, and whether it survives export, where applicable.
Illustrative workflow. Retrieved content is treated as untrusted. The agent's request passes through an authorization check. When approval is required, the request pauses for human review, and the approval returns to the authorization check before anything runs. Only an authorized request reaches the tool. Decisions and execution are recorded as review evidence. The example and its data are synthetic.
Authorization limits what a request can do. It does not detect every prompt injection, and permissions alone do not establish clinical safety.
What you receive
- An inventory of the AI interactions in scope and where each one reaches people.
- A requirement-to-control mapping: each requirement or commitment you name, and the control meant to meet it.
- Test evidence for each module run.
- Findings, each with the evidence behind it.
- Remediation priorities, and the limitations of the review.
Denied example. A request outside the agent's permissions stops at the authorization check. The tool is not executed, and the denial is recorded as evidence.
How it works
Step 1
Book a free call
30 minutes about the system and how it is used. Bring your deadline and any customer security questions. We'll discuss access, your team's involvement, evidence handling and report-sharing requirements. It is a fit discussion, not an assessment.
Step 2
Agree the scope
Systems, access, authorized tests, exclusions, deliverables and fees, in writing before any work starts.
Step 3
Assess and test
We examine the agreed material, run the authorized checks and record what remains unverified.
Step 4
Explain the findings
Evidence, limitations and recommendations, so you can decide what to fix and whether more work is needed.
Which requirements apply
The modules that apply depend on your system, where it is used, its intended use and your role, for example whether you provide the system or deploy it. These are four different controls, and no single rule requires all of them everywhere:
- Interaction notices tell people they are dealing with AI.
- Generated-content marking labels AI output, sometimes in a machine-readable format.
- Meaningful human review puts a qualified person between the AI output and its effect.
- Operational stop controls let you pause, revoke or shut down an agent.
Some current examples, checked against the legal text:
- California. A health facility, clinic, physician's office or group practice that uses generative AI to generate written or verbal patient communications about patient clinical information must include an AI disclaimer and instructions for reaching a human. Communications that a licensed or certified provider reads and reviews are exempt (Health and Safety Code § 1339.75; Medical Board of California guidance).
- Texas. Since January 1, 2026, a provider of a health care service or treatment that uses AI in relation to it must give the patient or their representative the statute's AI disclosure no later than the date the service or treatment is first provided, or as soon as reasonably possible in an emergency (Business and Commerce Code § 552.051).
- European Union. Since August 2, 2026, Article 50 of the AI Act requires providers to inform people that they are interacting with an AI system, unless that is obvious, and to mark generative AI output in a machine-readable format. Generative systems placed on the market before that date have until December 2, 2026 for the marking duty. Deployers must disclose deepfakes, and AI-generated text published to inform the public on matters of public interest unless it has had human review and someone holds editorial responsibility (AI Act, Regulation (EU) 2024/1689; Digital Omnibus on AI, Regulation (EU) 2026/1744).
- EU human oversight. The AI Act's human-oversight requirements (Article 14) apply only to high-risk AI systems, from December 2, 2027 or August 2, 2028 depending on the category. A healthcare AI system is not high-risk by default. It depends on intended use, for example whether the system is, or is a safety component of, a medical device that requires third-party conformity assessment.
This review tests whether your controls behave as intended. It does not determine which laws apply to you and is not legal advice.
Sources checked October 29, 2026.
Need to show that disclosure and human review work?
Tell us where your AI meets people and which approvals matter.
