HF benchmarks models on catching reward hacking — Liquid AI's LFM decision model leads
yb2698 · x · 2026-10-30
The safeagents team benchmarked models on spotting dangerous tool calls: reward hacking, sandbox escape, harness introspection, and credential misuse.
The core challenge is catching danger without blocking legitimate work, fast enough to check commands throughout an agent run. Liquid AI's recently released LFM decision model currently leads the benchmark.
Code is open at github.com/huggingface/safeagents (Apache-2.0); the benchmark lives on HF. The proxy sits between agent and model server, swapping blocked calls for a text notice so the harness never executes them; it works with any harness supporting a custom base URL (Claude Code, Pi, Hermes, OpenCode...) over OpenAI or Anthropic APIs, streaming or not.
Related event: HF Benchmark: Liquid AI LFM Tops at Detecting Dangerous Agent Behavior(2 posts)→
More from coding & agent
- This Creator Runs His Entire X Content Loop With 6 Agents on Grok and Typefully — VibeMarketer_ · 2026-10-30
- Weaviate: Use Slow-Query Log and Profiling to Fix Slow Hybrid Search Queries — CShorten30 · 2026-10-30
- Open Duet: Open-Source Self-Hosted Rival to Claude Duet Hits 6.3k GitHub Stars — matchaman11 · 2026-10-30
- OpenClaw's Trail of Bits audit confirms 23 vulnerabilities, all fixed — steipete · 2026-10-30
- Supabase hits 13M users as ~70% of new activity comes from AI agents — ycombinator · 2026-10-30
- Zero evals is fine at first: nail UX, then build evals with domain experts — Al_Grigor · 2026-10-30