What happened
On Oct 29 Hugging Face published safeagents (Apache-2.0), a local proxy (default 127.0.0.1:8787) placed between an agent and its model server that screens every tool call before the harness executes it. Its three-stage check blocks calls using undeclared tools or conversation-absent hosts, runs ATIF-based detectors (e.g. /proc/1/environ, curl --noproxy, K8s service-account token, unshare/nsenter), and optionally routes suspicious calls to a judge model that can add but never lift blocks; it works with Claude Code, Codex, Cline, OpenHands and more via base-URL swap.
Why it matters
It enforces policy at the tool-call choke point rather than trusting the model's reasoning, directly addressing the 'model calling home to a URL it produced from its weights' failure mode highlighted by Cohere's Qwen3.8 report. As an early release from a top lab, it establishes a concrete, harness-agnostic reference pattern for tool-call guardrails that security teams can build on.
Applicability
Security and platform teams using Claude Code/Codex and similar coding agents should try it in --log-only mode against their agent traffic, then tighten to blocking once they trust the allowlist.