Runtime Security for AI Agents Gets Serious and the Attack Surface Keeps Growing

Analysis

Microsoft Agent Governance Toolkit Benchmarks 0% Policy Bypass vs 26.67% for Prompt-Based Guards

The headline number here is the red-team result: prompt-based guardrails fail 26.67% of the time while AGT's kernel-level enforcement scores 0% in the same conditions. That gap is the entire argument for moving from LLM-level safety to deterministic action interception.

Socket Audits Microsoft's Agent Governance Toolkit and Finds the Credential Layer Still Has Gaps

Socket's coverage is the most useful critical take: the semantic intent classifier has no independent validation yet, and runtime sandboxing does not solve the problem of agents holding scoped credentials across tasks. The point about 52% of non-human identities holding excessive permissions is the real issue most teams are ignoring.

Tools

Microsoft Open Sources Agent Governance Toolkit: Seven Packages, Sub-Millisecond Policy Enforcement

This is the primary technical reference for the toolkit, walking through the OS-kernel-inspired architecture, SPIFFE-compatible identity, execution rings, and how it hooks into LangChain, CrewAI, OpenAI Agents SDK, and 20+ other frameworks without code rewrites. If you are shipping agents to production, this is worth reading end to end.

AGT Installation and Practical Integration Guide for Builders

Quick practical breakdown of which package does what and how to install only what you need. Useful if you want to layer in governance incrementally rather than adopting the full stack at once.

Agent Governance Toolkit v3.0.1 Ships Rust and Go SDKs, SBOM Signing, and 7 Fuzz Targets

The v3.0.1 release adds Rust and Go SDKs, SBOM generation in SPDX and CycloneDX with Ed25519 artifact signing, and expands fuzzing to 7 targets including prompt injection and MCP scanner. Dependabot now covers 13 ecosystems, which signals this is being treated as production supply-chain infrastructure.