Posted in Security Operations

Singapore just named the three tests an AI agent has to pass before anyone should trust it. Jasmin Lau…

Schomer — Principal Security Solutions Architect & Founder

Singapore just named the three tests an AI agent has to pass before anyone should trust it.

Jasmin Lau, Singapore's Senior Minister of State for Digital Development and Information, told Parliament on Nov 4: an agent earns trust through identity, authority, and accountability. Which agent is this, and who is responsible for it. What is it permitted to do, and on whose behalf. And can its actions be traced, so someone answers for them when it goes wrong.

Her line: "Agents can act for us, but they cannot answer for us."

Those are the same three checks the gates I build enforce — technically, on the wire, before the action happens:

Authority. The Modbus gate blocks the firmware write to [52,53] and forwards the legit write to [52,52]. The token gate blocks Bob withdrawing from Alice's account. The shell gate denies rm -rf /. The agent asks; the gate decides what it's allowed to do.

Accountability. Every decision lands in a record — allowed, denied, why. The STIX records now carry valid IDs and a declared extension, so a reviewer can actually trace a decision back to the rule that made it.

Identity. The honest gap. My HTTP gate checks a signature, but the key is a demo key — that's a placeholder, not an identity system. Singapore's answer here is a registry; mine is still being built.

The point isn't that my gates solve Singapore's problem. It's that the problem is the same one, and it's now being stated in Parliament: trust isn't a feeling about an agent. It's three checks, enforced before the action, with someone accountable after it.

7 reactions · 4 comments

Comments

Victor Font: Schomer, what I appreciate most is your willingness to identify the gap in your own architecture rather than claim the controls are more mature than they are. That's exactly the kind of transparency meaningful assurance requires. Identity, authority, and accountability are distinct requirements, and satisfying two doesn't compensate for the absence of the third. I'd add that accountability extends

Schomer: Appreciate that perspective, Victor. You nailed the exact boundary between technical enforcement and enterprise governance. A deterministic gate can enforce a rule and output a STIX record, but it can't invent the authority behind that rule. Bridging that gap—binding the cryptographic identity of the human who authorized the policy directly to the runtime telemetry—is the exact problem we're tackl

Victor Font: Schomer Schomer, that's an important next step, and I think you're approaching it from exactly the right direction. Binding an authenticated policy approval to runtime enforcement creates a traceable connection between human authority and machine execution. The additional challenge is preserving that chain when policies are modified, delegated, revoked, or superseded. A cryptographic signature can

Schomer: You're zeroing in on the exact problem that breaks most compliance models, Victor. A signature proves intent at time A, but without runtime state binding, it can't prove validity at time B when an agent actually executes. We address that by tying every deterministic gate decision to an immutable policy version hash inside the STIX output, while treating revocation as an immediate state-drop at the

Read and join the discussion on Cyberr