Kihan Sets Independent Authorization for AI Agents
The architecture checks agent actions outside the model, with enterprise evaluations targeted for December 2026 and production integrations still in development.
Updated on Nov. 5, 2026 in Artificial Intelligence
Live Poll
Would you trust a company's AI system to move your money without human approval?
San Francisco-based Kihan has launched Constitutional Computing, an architecture that places authorization for AI agents in an independent system outside the model. Its first product, Inktomi, has an operational decision engine, but production enforcement integrations are still under development, so the work is announced and at an early stage rather than broadly deployed.
Why it matters
As AI agents move from answering questions to executing transactions and changing production systems, the gap between what an agent can do and what it is permitted to do becomes a security problem. Kihan's approach, which checks each consequential action against signed organizational rules before execution, reflects a broader shift toward governance layers built apart from the models themselves.
The Cloud Security Alliance survey found 53% of organizations reported AI agents exceeding intended permissions and 47% reported a security incident tied to AI agents in the prior year. The source does not provide performance benchmarks for Inktomi's decision engine, and its enforcement integrations remain under development.
The players
Kihan
A San Francisco-based technology firm building enforceable authorization infrastructure for AI agents, with its first product Inktomi.
Cloud Security Alliance
A security research organization whose survey on AI agent permissions is cited as evidence of the problem Kihan addresses.
IEEE Computer
The publication of the Institute of Electrical and Electronics Engineers that is running AI governance feature articles in late 2026 and early 2027.
The details
Kihan's architecture moves authorization out of the AI model and into an independent decision engine that checks each proposed action against signed organizational rules. Permits are cryptographically verifiable, meaning they can be mathematically checked to confirm they were issued and not altered, and they are bound to the specific operation and its parameters. The tool that performs the action verifies the permit before acting, and the engine can permit, refuse, or refer a request for human approval, logging each decision in a tamper-evident record.
Timeline
April 2026: Cloud Security Alliance publishes its AI agent permissions research.
December 2026: Initial Inktomi enterprise evaluations are targeted, and the first of two IEEE Computer feature articles is scheduled.
January 2027: A second IEEE Computer feature on prompt framing is scheduled for publication.
The Tech Race
Kihan's approach of separating authorization from the model places it among efforts to build governance into agent infrastructure rather than relying on model behavior alone. The stub does not name specific competing products, so the competitive comparison is limited to this architectural choice.
Inktomi is designed to run on customers' own infrastructure, so organizations deploying AI agents would host the enforcement layer themselves. Live demonstration requests are being taken through Kihan's website, but production enforcement integrations are still under development, so the product is not yet broadly available.
The takeaway
The trajectory to watch is whether enforcement layers outside the model become standard infrastructure for AI agents. Track Kihan's initial enterprise evaluations, targeted for December 2026, as the first test of Inktomi's decision engine in real deployments.
What happens next
Initial Inktomi enterprise evaluations are targeted for December 2026, and two IEEE Computer feature articles are scheduled for December 2026 and January 2027.
Further reading
For the wider context of autonomous software systems, see Artificial Intelligence coverage.
More information
Request a live demonstration of Inktomi from Kihan.
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?
