AI Agents at RSAC 2026: What the Security Community Is Actually Watching

Analysis

Microsoft's Agent Governance Toolkit Gets Full Breakdown from Help Net Security

Help Net Security walked through all seven packages in detail, including the AgentMesh trust scoring system and the Agent Lightning RL governance module that most coverage has skipped. If you are evaluating whether AGT fits your stack, this is the most complete writeup outside of Microsoft's own blog.

Socket Flags the Credential Scoping Problem Runtime Governance Still Does Not Solve

Socket's writeup is the most skeptical take on AGT so far, pointing out that sandboxing limits what an agent does inside a process but does not stop a compromised agent from using credentials it already holds to hit downstream services. The gap between issuing short-lived credentials and actually revoking them per task is still a manual problem most teams are not solving.

Security

36.7% of MCP Servers Vulnerable to SSRF and 1,000 Have No Auth at All

BlueRock Security's analysis of 7,000+ MCP servers found more than a third are vulnerable to server-side request forgery that can leak AWS credentials from cloud metadata endpoints. That stat alone should be driving MCP security to the top of your backlog if you are running agents against external tools.

News

68% of Organizations Cannot Distinguish Human from AI Agent Activity in Their Logs

This stat from a CSA survey of 228 IT and security professionals buried in the AGT launch post is the real story for anyone running agents in production. Only 18% are confident their IAM systems can handle agent identities at all, which means most teams are flying blind on attribution before they even get to policy enforcement.

Tools

AGT v3.0.1 Ships SBOM Generation with Ed25519 Signing and Expands Fuzzing to 7 Targets

The v3.0.1 release added SPDX and CycloneDX SBOM output with cryptographic signing, which matters for any team that needs supply chain evidence for EU AI Act or SOC 2 compliance. Dependabot now covers 13 ecosystems including Cargo, Go modules, NuGet, and Docker, which is a meaningful step toward keeping the governance layer itself from becoming an attack surface.