EU AI Act Logging Rules, MCP Server Exposure, and the Governance Gap Widens

Analysis

What the EU AI Act Actually Requires for AI Agent Logging

The EU AI Act's high-risk AI obligations land in August 2026 and logging is one of the concrete technical requirements teams will need to implement now. If you're shipping agents into regulated environments, this is the spec you need to read before your compliance team does.

Socket's Deep Dive on the Limits of Microsoft's Agent Governance Toolkit

Socket does a honest teardown of where runtime policy enforcement actually breaks down, specifically around credential scoping and per-task permission revocation that most orchestration frameworks still don't support natively. Worth reading before you assume a governance layer solves your IAM problem.

Security

Bitsight Finds ~1,000 Exposed MCP Servers With No Authorization

Roughly 1,000 MCP servers are sitting on the open internet with zero authorization, and BlueRock Security found 36.7% of 7,000 analyzed servers vulnerable to SSRF attacks that can pull cloud credentials. The attack surface for agent tool integrations is much larger than most teams realize.

Fortinet Patches Critical FortiSandbox Vulnerabilities Affecting Agent-Adjacent Infrastructure

Critical patches for FortiSandbox land the same week agent sandbox architectures are getting serious attention, a reminder that the underlying security infrastructure agents run on top of needs the same scrutiny. Patch your sandboxing layer before you govern what runs inside it.

Tools

Agent Governance Toolkit v3.0.1 Adds Rust and Go SDKs Plus SBOM Signing

The toolkit now ships across five languages with SBOM generation in SPDX and CycloneDX formats, Ed25519 artifact signing, and 7 fuzz targets covering prompt injection and MCP scanner paths. Incremental adoption just got easier for teams not running Python stacks.

Microsoft's Agent Governance Toolkit: The OS-Kernel Mental Model for Agent Security

The framing here is what makes this worth revisiting: treating agent governance like an OS kernel, with privilege rings, stateless policy enforcement, and circuit breakers borrowed from SRE, rather than bolting on prompt guardrails. Sub-millisecond policy evaluation at 9,300 ops/sec means governance overhead is no longer an excuse to skip it.