Kihan launches Constitutional Computing to authorise AI agent actions before they run

Kihan, a new San Francisco company, launched on 5 November with what it calls Constitutional Computing, an architecture that separates an AI agent's ability to act from its authority to do so. According to the company, each consequential action is checked against signed rules outside the AI before it runs, and the executing tool must verify a cryptographic permit bound to that exact operation. Actions can be permitted, refused or referred for human approval, with a tamper evident record of each decision.

Kihan's first product, Inktomi, targets agents built on models from Anthropic, OpenAI and Google and tools that use the Model Context Protocol. The company says its decision engine already works, while production enforcement integrations are still in development, with first enterprise evaluations targeted for December 2026. Inktomi runs on customers' own infrastructure and is aimed first at financial services, insurance, healthcare, legal services and other regulated industries.

Kihan's release cites Cloud Security Alliance research from April 2026 in which 53% of surveyed organizations reported AI agents exceeding their intended permissions and 47% reported an AI agent security incident in the previous year.

"Telling an AI to behave itself isn't security," said co-founder and CEO Jonathan Hirshon. The company says the research behind the architecture will appear as a two part feature in IEEE Computer, starting with the magazine's December 2026 AI Governance issue. More details are on Kihan's website.