Hugging Face Ships atif-scan 0.8.16 to Flag Reward Hacking in Agent Traces
yb2698 · x · 2026-11-04
Hugging Face released atif-scan 0.8.16, a tool for checking benchmark traces for reward hacking, time skewing, awareness and more.
- Adds an improved MCP Server for trace analysis plus better checks and handling for Kimi Code and tb4 tasks/allowances.
- Uses plugin-style detectors and simple rules over ATIF agent trajectories; the analysis core is offline, standard-library Python, with the CLI adding huggingfacehub (Hub inputs only) and rich.
- It never runs anything found in a trace and never calls a model unless --image-model is passed.
- Scans local directories or downloads and caches from Hugging Face buckets or Harbor; can read mirrors of Datacurve's public DeepSWE trials with their trial index.
- Findings are review candidates, not verdicts; severity is a review priority, not a probability of cheating, and unknown evidence is never treated as clean.
More from coding & agent
- 8x Microsoft MVP ships complete Microsoft Agent Framework series and enterprise agents course — marlene_zw · 2026-11-04
- Living Brain tracks 3 health scores to stop AI agents relying on stale memory — cneuralnetwork · 2026-11-04
- Open-source skill makes Claude Code ship polished, notarized macOS apps in one shot — vista8 · 2026-11-04
- Power user tallies 12B coding-agent tokens, finds 96% were cache reads — peakpirate007 · 2026-11-04
- Indian code-mixed calls have no single correct transcript: LLM, CRM and compliance each need a different one — BEYOND__7 · 2026-11-04
- Alexey Grigorev releases free kit of 28 prompts to ship an app from spec to production — Al_Grigor · 2026-11-04