Kihan launches constitutional computing for enterprise AI agents

Kihan, a new San Francisco company, said Wednesday it is building Constitutional Computing, an architecture meant to stop unauthorized AI agent actions before they reach enterprise systems. The company also said its research will appear in a two-part IEEE Computer feature starting with the magazine’s December 2026 AI Governance issue.

Why it matters: - Enterprise AI agents are moving beyond chat and into actions that can move money, change systems and access sensitive data. - Kihan is targeting the gap between what an AI agent can do and what it is allowed to do. - The company’s pitch is that governance should be enforced before execution, not after a mistake has already affected production systems. - A Cloud Security Alliance report from April 2026 found 53% of surveyed organizations said AI agents exceeded intended permissions, and 47% reported an AI-agent security incident in the prior year.

What happened: - Kihan announced its formation in San Francisco on November 5, 2026. - The company introduced Constitutional Computing, a mathematically based architecture designed to make AI governance enforceable at the point of execution. - Kihan said the architecture is built to prevent unauthorized AI agent actions before they can affect enterprise systems. - The company said its foundational research will appear in IEEE Computer, beginning with the magazine’s December 2026 special issue on AI Governance.

The details: - Kihan’s system separates AI capability from permission to act. - Authority is written, signed and enforced outside the AI model. - Each consequential action is checked against governing authority before execution. - The system uses a cryptographically verifiable permit tied to the specific operation and its parameters. - The executing tool verifies the permit before acting. - On protected execution paths, no valid permit means no execution. - The architecture creates tamper-evident records showing what was authorized, refused or escalated for human approval. - Kihan’s first product, Inktomi, is being developed for enterprise AI environments built on models from Anthropic, OpenAI and Google. - Inktomi is also being built for agent frameworks such as Hermes Agent and tool ecosystems using the Model Context Protocol, or MCP. - Kihan said Inktomi’s decision engine is already operational. - The decision engine evaluates specific actions against signed rules and can permit, refuse or escalate a request. - Production enforcement integrations are still under development. - The company is targeting initial enterprise evaluations for December 2026. - Inktomi is designed to run on customers’ own infrastructure. - Kihan said the product is meant to complement existing identity management, cybersecurity and AI governance tools rather than replace them. - The company is initially targeting financial services, insurance, healthcare, legal services and other regulated industries. - Kihan is exploring integration with the NVIDIA OpenShell runtime and the IBM watsonx.governance platform. - The first IEEE Computer article is titled "Checking LLM Outputs Before They Become State: The Problem and the Admission Gate." - The second article is titled "Checking LLM Outputs Before They Become State: Prompt Framing Across Four Models." - The first article is scheduled for the December 2026 issue, followed by the second in January 2027.

Between the lines: - Kihan is betting that enterprise AI will need independent authorization controls, not just model instructions or policy prompts. - The company is framing governance as an infrastructure problem, which could make it more relevant to regulated buyers than a model-tuning or prompt-safety pitch. - By emphasizing signed rules, verifiable permits and execution-time enforcement, Kihan is positioning itself closer to security infrastructure than to traditional AI tooling. - The IEEE Computer publication adds third-party credibility to the underlying approach and may help Kihan reach technical decision-makers.

What's next: - Kihan plans live demonstrations of Inktomi for prospective customers. - The company’s initial enterprise evaluations are slated for December 2026. - Production enforcement integrations will determine how quickly the architecture can move from research to deployment. - The IEEE Computer feature will roll out across December 2026 and January 2027.

The bottom line: - Kihan is trying to turn AI governance from a policy promise into an execution-time control layer for enterprise agents.