Wisevoter
U.S.

Kihan Launches Constitutional Computing for AI Agents

San Francisco firm Kihan introduced an architecture that places authorization for AI agents in an independent system outside the model.

Updated on Nov. 5, 2026 in Artificial Intelligence

Isometric editorial illustration of three stacked solid blocks with a brass key-shaped gate passing between them, in oxblood, teal, mustard, and cream.
Kihan launched Constitutional Computing, an architecture that moves authorization for AI agents into an independent system outside the model, from its San Francisco base. AI Illustration. Upload story photo >

Live Poll

Would you trust a company's AI system to move your money without human approval?

Kihan, a San Francisco-based technology company, launched Constitutional Computing, an architecture for enforceable AI agent authorization. The company says the approach addresses the gap between an AI agent's ability to act and its authority to act.

Why it matters

AI agents increasingly execute transactions, modify production systems, and access sensitive corporate information. Research by the Cloud Security Alliance found 53% of surveyed organizations reported AI agents exceeding their intended permissions.

Each consequential action is evaluated against governing authority before execution, and permits are cryptographically verifiable and bound to the specific operation and its parameters. Production enforcement integrations for Kihan's first product, Inktomi, are still under development.

The players

Kihan

Kihan is a San Francisco-based technology firm founded to close the gap between what AI agents can do and what they are authorized to do.

Cloud Security Alliance

The Cloud Security Alliance is a cloud security industry organization whose research found 53% of surveyed organizations reported AI agents exceeding their intended permissions.

The details

An independent decision engine checks an agent's proposed action against signed organizational rules, and requests can be permitted, refused, or referred for human approval with tamper-evident records of each decision. The executing tool verifies the permit before acting, and Inktomi is designed to run on customers' own infrastructure.

Timeline

  1. April 2026: Cloud Security Alliance publishes AI agent permissions research.

  2. December 2026: Initial Inktomi enterprise evaluations are targeted, and an IEEE Computer AI Governance special issue is scheduled.

  3. January 2027: A second IEEE Computer feature on prompt framing is scheduled for publication.

Roadmap

Kihan's launch responds to a documented pattern in which AI agents operate beyond their intended permissions, a trend the Cloud Security Alliance quantified in its research. The approach moves authorization outside the model itself, a design choice that shifts governance from the AI system to an independent enforcement layer.

Organizations deploying AI agents that execute transactions or modify production systems gain a way to enforce limits on what those agents can do, rather than relying on the models to stay within bounds. Enterprises evaluating the technology would be able to run the decision engine on their own infrastructure once it is available to them.

The takeaway

Organizations adopting AI agents should treat authorization as a separate control from the model's own behavior. Verifying permissions before each consequential action is a practical safeguard against agents exceeding their intended scope.

What happens next

Initial enterprise evaluations of Inktomi are targeted for December 2026, and two-part IEEE Computer feature articles are scheduled for December 2026 and January 2027.

Further reading

For more on the technology behind this launch, explore Artificial Intelligence coverage on Wisevoter.

More information

Request a live demonstration of Inktomi through the Kihan website.

Source note: This article includes information reported by GamesBeat.

Live Poll

Would you trust a company's AI system to move your money without human approval?