Microsoft Agent Governance Toolkit vs OpenAI Privacy Filter

The Microsoft Agent Governance Toolkit is an open-source runtime security and policy enforcement framework for autonomous AI agents. It covers all 10 risks in the OWASP Agentic AI Top 10 — from prompt injection and excessive agency to memory poisoning and supply chain vulnerabilities. The toolkit provides sub-millisecond policy hooks that integrate with LangChain, CrewAI, Google ADK, and most other major agent frameworks, across Python, Rust, TypeScript, Go, and .NET. The core approach is "policy as guardrail": rather than trying to make agents safe by constraining their prompts, the toolkit enforces runtime boundaries on what agents can actually do — file access, API calls, tool invocations — before execution happens. Think of it as a capability firewall for agents, similar to how AppArmor works for Linux processes. As enterprises push AI agents into production, governance and compliance are becoming blockers. The toolkit was designed in collaboration with Microsoft's security research teams who've been auditing internal agentic deployments. It ships with a policy library covering common enterprise scenarios (PII access, external API calls, sensitive file paths) and a dashboard for audit logging — addressing the 'how do I explain what my agents did' problem that's stalling adoption in regulated industries.

OpenAI released Privacy Filter on April 22, 2026 — a 1.5B-parameter open-weight model for detecting and redacting personally identifiable information from text before it ever reaches a cloud API. The model runs fully locally, handles 128,000 tokens in a single pass, and achieves a 96% F1 score across eight PII categories: names, addresses, emails, phone numbers, URLs, dates, account numbers, and secrets. Unlike traditional regex-based PII scrubbers that choke on unstructured text and context-dependent references, Privacy Filter uses a fine-tuned language model to understand semantic context — it catches "call me at the usual number" type references that pattern matchers miss entirely. The model ships with only 50M active parameters at inference time via sparse activation, keeping latency low enough for preprocessing pipelines. Available on Hugging Face and GitHub under Apache 2.0, Privacy Filter solves a real bottleneck: enterprises and regulated industries have been unable to safely pipe sensitive documents through LLMs at scale. OpenAI explicitly warns it should be treated as a "redaction aid, not a safety guarantee," which is unusually honest for a model card — and a sensible framing for high-stakes medical or legal workflows.