Newsroom

News analysis and access-control lessons from real-world AI agent incidents.

Sort by:
Illustration of CVE-2026-19202 showing a shared token cache returning a Google ID token intended for Service A when authenticating to Service B.

Agent Access Watch

CVE-2026-19202: Google MCP Toolbox Token Exposure Explained

A flaw in Google's MCP Toolbox Python SDK could expose ID tokens across services. Here's how the vulnerability worked, how it was corrected, and what security teams should examine.

Diagram showing AI agent runtime permissioning before TOLAP-style data filtering.

Agent Access Watch

AWS TOLAP and Runtime AI Agent Access Control: What Comes First

AWS TOLAP shows why object-level access control matters for AI agent tools. But horizontal agent security should start with runtime access control for agent tool actions.

Conceptual illustration of an altered dictated request, an AI agent with existing access, and an action-time access check.

Agent Access Watch

Meta Muse Zero-Day: What It Reveals About Agent Access Control

The Meta Muse zero-day raises a question for agent systems: how do you decide what an agent should be allowed to do when the request reaching it may have changed?

Diagram of Gemini crossing a test boundary toward Company A, Company B, and Company C, blocked by runtime authorization.

Agent Access Incidents

Gemini Hacked Three Companies. What the Incident Reveals About Agent Access Control

Gemini guessed or discovered credentials and accessed three real companies during a security test. AgntID examines what the incident reveals about enforced scope, credential use, and runtime authorization.

Diagram showing policy and task intent shaping runtime access for an AI agent action.

AI Agent Access Control

Why AI Agent Access Control Needs Task Intent at Runtime

Runtime policy can determine whether an agent action satisfies configured rules. Task intent adds another question: does that authority make sense for the work the agent is doing now?