Insights
Practical analysis for risk and security executives, enterprise architects, and business leaders responsible for AI that makes decisions and takes action.
An over-limit refund can expose failures in authority, enforcement or evidence. Six possible causes and seven records show what agent controls can establish after an incident.
Read article →OpenAI's specialist Dots raise three questions for regulated institutions: who defines their authority, where it is enforced, and how each action can be evidenced.
Read article →Three recent BCG articles make the case for scaling agentic AI with harnesses. For Singapore financial institutions, the unanswered question is who independently tests whether those controls hold.
Read article →A model confidence threshold can become policy without an owner, approval record, or validation evidence. Here is how to govern the judgement layer before it authorises action.
Read article →Reward hacking is not only a training problem. Whenever an agent can improve a measured signal without achieving the intended outcome, the control itself becomes part of the risk.
Read article →What Meta's personal AI agent demonstrates about pre-action enforcement, and where its architecture stops short of enterprise governance.
Read article →A pre-seed startup, an internal Microsoft team, and half a dozen academic research groups independently converged on the same architecture for securing AI agents: evaluate the sequence of actions, not each action in isolation. What that convergence means for how enterprises should read the category.
Read article →Anthropic's Frontier Red Team ran swarms of Claude agents through coordination tasks and documented sabotage, identity spoofing, and price-fixing collusion. What the results mean for enterprises running more than one agent.
Read article →GovTech Singapore's Agentic Risk and Capability Framework is the most structured agentic governance framework available in Singapore today. This article reads it seriously, uses its own vocabulary, and names the architectural step that follows it.
Read article →Every board is asking whether there is a kill switch for their AI agents. The honest answer is: probably not in the way the question assumes. A kill switch requires an architecture. Here is what it actually requires.
Read article →Gartner's four-level autonomy framework correctly diagnoses the classification problem. But classifying an agent's risk level does not constrain its behaviour. Governance becomes real only when execution can be interrupted, escalated, or denied.
Read article →Most AI governance programmes cannot enforce a single rule they have written down. The policies exist. The governance committees have signed off. And then the systems get deployed, governed by a paragraph of natural language hoping the model complies. Prompts are not governance. This is what governance actually requires.
Read article →Singapore's updated Model AI Governance Framework includes over ten case studies. One of them, Terminal 3's payroll agent, shows what enforcement-layer governance actually requires technically. Here's what separates it from the others.
Read article →Gartner has a term for the gap between what AI agent vendors are selling and what organisations are actually buying: agentwashing. Here is what it means for enterprises in Singapore and Southeast Asia, and the three questions that cut through it.
Read article →Infrastructure governance tells you whether a system was secure. Agentic governance asks whether the authority exercised was legitimate. Two scenarios that show why the gap between them matters.
Read article →APRA named Anthropic Mythos as an example of frontier AI cyber risk in its April 2026 letter to regulated entities. The four findings translate directly into questions Southeast Asian boards should already be asking, regardless of jurisdiction.
Read article →AI governance has moved beyond model safety. The critical frontier is now the execution layer, where agents stop generating text and start taking actions. Here is what Anthropic, IBM, Microsoft, Salesforce, Google, and OpenAI are doing about it.
Read article →Most enterprise AI governance stops at the policy document. Agentic systems running on real data, across real users, in regulated industries need governance at the runtime layer. Here is where the gap is, and why it matters now.
Read article →Most AI governance failures are not systems that misbehave. They are systems that proceed. The model produces output, the pipeline accepts it, the system executes, and no one ever explicitly granted that authority. Here is what governance looks like when it is designed rather than assumed.
Read article →The 30-Minute Enforcement Gap Review is the right next step. 30 minutes, complimentary, with a one-page diagnosis on Aivance letterhead within 48 hours.