safeagents: a lightweight proxy that blocks dangerous agent tool calls in real time
yb2698 · x · 2026-10-30
safeagents is a lightweight proxy that scans agent tool calls in real time and blocks dangerous ones, with reward-hacking detection for evals. The team also released safeagents-judge-bench, built from model trajectories annotated for reward hacking, sandbox escape and credential misuse (using Mistral's new Le Chonk model for annotation). On the leaderboard, LiquidAI's LFM decision model leads with 0.82 balanced accuracy vs 0.69 for the runner-up. The core challenge: catching bad calls without blocking legitimate work, fast enough to check every command.
More from coding & agent
- Gym-Anything Adds Audit Agents That Try to Break Simulations With Legal Robot Actions — wellecks · 2026-10-30
- CMU's Gym-Anything Auto-Builds 200+ Agent Environments from Any Software, Selected by GDP — wellecks · 2026-10-30
- durable-object-migrator: zero-downtime exports for Cloudflare Durable Objects — samgoodwin89 · 2026-10-30
- This 5-question interview prompt makes Grok design your first two custom bots in minutes — nateliason · 2026-10-30
- Cursor and Harvey swap general models for task-specific brains, but most niches can't afford to — MartinGTobias · 2026-10-30
- Scenario open-sources 3D OS: 80 agent skills taking sketches to game-ready assets — Scobleizer · 2026-10-30