Agent Armor vs OpenAI Privacy Filter

Agent Armor is a lightweight governance layer for AI agents, written in Rust and designed to intercept every agent action before execution. It sits in front of LangChain, CrewAI, AutoGen, or Claude Code and runs each proposed action through an 8-stage decision pipeline: intent classification, credential leak scanning, rate limiting, resource scoping, behavioral fingerprinting, semantic deduplication, human-review escalation, and final allow/block. The project is MCP-aware and can intercept tool calls at the protocol level, which means it works regardless of which agent framework you're using. Actions that pass all 8 layers execute normally; those that fail can be automatically blocked, held for human review, or rewritten to a safer equivalent. A live dashboard shows agent activity, pending reviews, and anomaly alerts. Version 0.3.0 arrived as a Show HN today and hit the front page. The author, Edoardo Bambini, built it after a production incident where a coding agent attempted to overwrite git history on the main branch. The timing is good — as more teams ship agents to production, "what guardrails do I put between the agent and the real world?" is an increasingly urgent question.

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.