Steven Willmott is chief executive of Safe Intelligence, an Imperial College London spinout developing tools to verify that AI systems operate safely in high-stakes environments. As co-founder and former chief executive of 3scale, he helped build the infrastructure and standards behind modern APIs; today, he is bringing the same insistence on explicit specifications to autonomous AI agents.
At 3scale, Willmott developed tools for managing the interfaces through which software systems communicate. Red Hat acquired the company in 2016, after which he became a senior director overseeing API management and infrastructure. He contributed to the OpenAPI Initiative, API Commons and APIs.json, and supported the release of OpenAPI 3.0 while at Red Hat.
He subsequently co-founded Timewarp, exploring AI and games, and argued in writing about generative AI that the technology’s practical value lies in what people can accomplish alongside it. At Safe Intelligence, founded by Imperial computer scientist Alessio Lomuscio, he leads a company that raised £4.15 million in seed funding to advance validation for safety-critical applications, including aviation and financial services.
The company initially applied formal verification and robustness testing to vision and tabular models, examining whether changes in operating conditions could produce dangerous behavior. Willmott has extended that approach to language-model applications through Spec27, a platform for validating AI agents against explicit behavioral requirements.
What reliable agents actually require
- Behavioral contracts, not benchmark scores. Effective specification-driven agent validation must capture business rules, user permissions, permitted actions and domain-specific terminology. A support agent should obey refund limits; an airline assistant should recognize which destinations its airline actually serves.
- Capability can expand the attack surface. Larger models may understand concealed malicious instructions that weaker systems miss, while access to financial or operational tools increases the consequences of failure. Safety depends on matching an agent’s capabilities and permissions to its actual responsibilities.
- Robustness under realistic variation. Reliable agents must withstand misspellings, rephrased requests and specialized vocabulary. Financial terms such as gross sales and gross profit, for example, cannot be treated as interchangeable merely because a language model considers them similar.
- Specifications that survive infrastructure changes. Willmott advocates implementation-independent behavioral specifications that can be version-controlled and reused when organizations replace models, frameworks or testing tools. His AI Engineer Europe presentation connects these durable contracts to security testing, regression checks and the practical challenge of defining what an autonomous system must never do.
