Three launches this week show how the problem of trusting AI agents is breaking into separate questions. Arkose Labs is trying to identify the agent itself, even when it hides. Sardine wants to show merchants which person is behind an agent’s purchase. And Kihan, a new San Francisco company, is focused on whether an agent has the authority to take an action before it takes it.
Arkose’s update starts from the finding that agents rarely announce themselves: its latest report says personal AI agents identify themselves less than 3 percent of the time. On Thursday the company expanded its Agent Trust Manager with Agent ID, a stable identifier that it says holds when an agent rotates networks, ships a new build or appears across many accounts. Each session is classified as self-disclosing good, non-disclosing good or malicious, and businesses can set a policy to allow, monitor or block agents at each step of a customer journey, such as sign-up, login, checkout and money movement. “Agents are already changing how they identify themselves to avoid being blocked, and they can pass authentication steps built for people,” said Kevin Gosschalk, Arkose’s founder and CEO. A Proof of Human feature, listed as coming soon, would require a person to confirm before an agent can continue at a sensitive step.
Sardine’s Know Your Human protocol looks at that person from the merchant’s side. Most agent-led shopping takes place in cloud browsers, the company says, so orders arrive from data center IP addresses with user agent strings associated with automation, and fraud systems decline them as bots. Under the protocol, an agent adds a few HTTP headers that identify a provider and session, and the merchant uses them through Sardine to retrieve the customer’s device fingerprint and IP address at checkout. A merchant that detects an agent checkout can also text the customer a verification link, valid for five minutes, that ties the customer’s browser to the agent’s cloud browser. Sardine also pitches the data as evidence in chargeback disputes, while saying that collecting a signal alone does not guarantee a successful response, and its proposal describes possible extensions to email, phone, KYC and identity verification data.
Kihan, which came out of stealth on Thursday, is concerned less with who an agent is than with what it may do. Its architecture, called Constitutional Computing, checks each consequential action against governing authority outside the AI, issues a cryptographically verifiable permit bound to that operation and its parameters, and has the executing tool verify the permit before acting. On protected paths, no valid permit means no execution. “Telling an AI to behave itself isn’t security. Authorization has to be independent of the AI, verifiable and enforceable before anything consequential happens,” said co-founder and CEO Jonathan Hirshon. Kihan says its first product, Inktomi, is meant to complement existing identity management tools rather than replace them.
Each approach answers a question the others leave open. A stable agent identifier does not say who sent the agent. A person’s device signals do not say what that person authorized. A permit works only where the tool checks for one. Those gaps run through other developments this week, in which LexisNexis Risk Solutions, Zetrix AI and U.S. bank regulators each took up what should be verified when an agent acts for someone.
Other vendors have tried to bundle the questions together. Shufti’s Know Your Agent, launched last month, checks an agent’s identity, confirms that a verified person or business authorized its action and enforces that owner’s permissions.
All three of this week’s launches are at an early stage. Sardine has built its protocol into its own Personal Agent and is seeking design partners among agent developers and merchants. Arkose’s proof-of-human step has not shipped. Kihan’s decision engine is running, but its production integrations are still in development, and its first enterprise evaluations are targeted for December.
Sources: Arkose Labs, Sardine, The Paypers, Kihan
–
By the ID Tech Editorial Team
