The idea
Most AI safety and security work looks at the model: what it was trained on, what it says, whether it can be made to say something it shouldn't. That work matters, but it stops one step short of where harm happens. Harm happens when an output becomes an action — a transfer is sent, a record is written, a query is run against a surveillance database, a payment is released on the strength of a photograph — and an action is only as safe as the check that stood in front of it.
SentinelPoT is a company about that check. The platform reads what a proposed action would do from the action's own parameters, spends validation effort in proportion to the consequence, decides before anything executes, and records a signed attestation whether the answer was yes, no, or ask a human. We call the method Proof-of-Trust™, and we apply it wherever an AI system, an agent, or a person acting through an AI interface is about to do something that is hard to take back.
Two convictions run through everything on this site. The first is that evidence beats assurance: a signed record that a check ran is worth more than any claim that a system is safe, so we build the record first. The second is that governance has to be affordable enough to leave switched on, which is why validation cost scales with what an action could do rather than being applied uniformly and then quietly disabled.
Founder
Chris Smith founded SentinelPoT and designed the Proof-of-Trust method and the Sentinel Runtime engine. His work sits at the intersection of AI security, runtime assurance and verifiable evidence: consequence-proportional validation for tool-using agents, tamper-evident attestation for automated decisions, and use-policy enforcement for systems that act on people. He is pursuing doctoral research in this area alongside the company, and Sentinel Labs is where that research is published.
He built the first version of the runtime to answer a practical question — how to let an agent hold real authority over money and infrastructure without either trusting it blindly or slowing it to uselessness — and the answer became the company.
How we work
We are small and deliberate about it. Every capability on this site is delivered through a scoped pilot: one system, four to six weeks, our team configuring and operating the gate with yours, signed evidence at the end. We would rather run three pilots well than sell a platform nobody has installed.
We publish what we can. Technical reports go out through Sentinel Labs with limitations sections and reproducible methods; the verifier for our attestations is open source so that no one needs our permission to check our evidence; and the Trust page lists the claims we make and the ones we refuse to.
We do not name customers, researchers or partners without their written consent, and we treat findings from every engagement as the customer's confidential information.
Contact
Pilots and partnerships: request a pilot or write to info@sentinelpot.ai. Security reports: security@sentinelpot.ai, under our disclosure policy. Researcher program: research@sentinelpot.ai. Corrections to anything we have published: labs@sentinelpot.ai.