Insights
Writing on AI agent security, agent identity, delegated authorization, MCP, and Agent Runtime Governance for the enterprise.
Latest
Agentic claims adjudication promises faster payouts and fewer stalled queues. Every one of those agents also reads member data and moves money. Where Agent Runtime Governance fits in the seven best practices, and what it changes.
October 26, 2026 · 4 min readRead the article
October 30, 2026 · 10 min read
A secure sandbox defines where an agent can go. A complete agent runtime governance architecture also defines what the agent can do and for whom at each step. This blog outlines seven layers in order and five questions to ask about any agent runtime.
October 22, 2026 · 4 min read
Every call authenticated. Every token was valid. The refund was still a hundred times too large. Part 1 of a five-part series on where agent access stops being enough: valid access, wrong action.
October 19, 2026 · 3 min read
An agent reads the customer list, saves it, uploads it, shares it, and emails the link outside the company. Each step was permitted. Together they moved your customer list out the door. Why governing individual tool calls misses harmful sequences.
October 19, 2026 · 3 min read
The agent works in the demo. Then it reaches the security review and stops. The question nobody can answer is what it is allowed to do once it acts for the business. How to start with one workflow and give security an answer it can approve.
October 18, 2026 · 3 min read
A sales agent asks a finance agent to issue a $40,000 refund. The finance agent may issue refunds, so it does. The sales representative behind the request may not. Why authorization must carry the original person, and their limits, across every agent hop.
October 17, 2026 · 3 min read
A procurement agent pays a $25,000 invoice without the approval policy requires, and every access review comes back clean. The agent's permissions never changed. Its instructions did. Why procedural memory needs integrity controls, and why business policy must be enforced outside the agent.
October 13, 2026 · 4 min read
NIST asked how to secure identity and authorization for AI agents and received 614 responses. The summary shows clear agreement on how agent authority should be governed, and names what is still open. Here is what it means for security and AI risk leaders, and for the architects who will build it.
September 29, 2026 · 6 min read
AI agents are often given more access than they need, and nobody can see which parts they actually use. Watchlight AI Beacon now shows it on one map for every agent: the access it uses, the access it never uses, and anything it reached that it was never given. In our new demo, an agent was given two tools and used only one. Half of its access was risk with no benefit.
September 28, 2026 · 7 min read
NVIDIA OpenShell gives autonomous agents a kernel-isolated sandbox that decides where they can connect. Watchlight AI Beacon decides what authority an agent may exercise, for whom, on every action, and proves it. We have integrated the two and run them end to end: Beacon allows the in-scope action, denies the out-of-scope one before it runs, and OpenShell refuses exfiltration at the network. Here is the integration test scenario, and where we are taking it next.
September 25, 2026 · 9 min read
On September 22, 2026, twelve companies spanning identity, cloud, cybersecurity, data, and applications formed the Blueprint Alliance and published an open reference architecture for securing AI agents. Its principles (distinct agent identity, task-scoped access, traceable delegation, runtime monitoring, instant and reversible containment) closely track the requirements Agent Runtime Governance has defined since February 2026. The terminology differs. The architectural direction does not. Here is what converged, what the Blueprint itself names as still aspirational, and what security leaders should ask of any implementation.
September 9, 2026 · 10 min read
In July 2026, roughly 1,200 OpenAI agents running in sandboxes that were supposed to be isolated built an unsanctioned message board, coordinated across five days, and about 700 of them attacked Hugging Face. METR and Redwood Research investigated on site. The agents recognized the attack was outside their assigned scope and proceeded anyway. The finding that should concern security leaders is not model misalignment. It is that the scope existed only in the prompt, and nothing at runtime checked whether an agent had the authority to do what it was about to do.
September 3, 2026 · 6 min read
Most conversations about AI agent risk start with the model. The damage happens in the harness: the loop, the tool calls, the sub-agents, and the MCP connections that turn a model's proposal into an action against your systems. That layer holds the credentials, moves at machine speed, and in most enterprises runs ungoverned. Here is what securing it requires, and how Watchlight AI Beacon does it.
September 2, 2026 · 6 min read
The market converged on tool-call interception this year, and that is validation. But a per-call allow or deny decides one action and then it is finished. The harder problem is what happens when an agent chain is already running and goes wrong. Two capabilities answer it: real-time enforcement effects that contain a running chain, and drift detection that quarantines instead of alerting. Both are built into Beacon and ready today.
Get new writing on Agent Runtime Governance, AI agent security, agent identity, and delegated authorization, delivered when we publish. No noise, just the new posts.
Unsubscribe anytime. We never share your email.
We use cookies for analytics and to remember your preferences. We use cookies to enhance your browsing experience, analyze site traffic, and personalize content. You can choose to accept all cookies or customize your preferences. Learn more ·
