AI agents are a new kind of identity. Treat them like one.
Agents act with credentials, reach real systems and persist across sessions. Identity and access management, not prompt wording, is where their risk is controlled.
Insights
Notes on deploying AI safely and well
Practical analysis for leaders and technical teams on agents, evaluation, security and governance, written from the questions organisations are working through now.
6 articles
Prompt injection is an architecture problem, not a prompt problem
Any system that reads untrusted content and can take actions is exposed. The defences that hold are structural: separation, scoped tools and confirmation outside the model.
Evaluating agents, not answers
A correct final answer can hide an unsafe path. Agent evaluation has to look at trajectories, tool use and cost, and it has to keep running after launch.
A sandbox is a claim, not a fact
Reported 2026 incidents in which evaluated agents reached real systems show why containment has to be verified, not assumed. The lesson applies well beyond frontier labs.
Human oversight that can actually change an outcome
A review step only counts as oversight if the reviewer has the time, information and authority to say no. Most designs fail on at least one of the three.
Why governance decides which AI projects reach production
Pilots stall less often on model quality than on unanswered questions about data, risk and ownership. Settling those early is what moves a project into operation.
No articles in this category yet. Select another category.
Each article sets out how we think about a question organisations are facing now. If there is a subject you would like us to cover, tell us.
Next step
Ask the question you would want an article to answer.
A short conversation is usually faster than waiting for published material. Describe the decision in front of you and we will respond with a considered technical view.