HF safeagents adds scrub & publish commands to redact agent traces before upload
yb2698 · x · 2026-11-06
Hugging Face's safeagents project adds agent trajectory scrubbing and publishing: new envs are built from community-shared traces on the Hub, and safeagents scrub / safeagents publish mask API keys, tokens, auth headers, private keys, secret env values, emails, home directories and IPs before rechecking and uploading redacted copies. Training streams live with reward curves, held-out pass@1 per checkpoint, and tool calls per solved task; in a 50-step ablation (LFM2.5-24B), enabling the guard changed held-out accuracy by +0.0 ± 4.1 points — essentially free insurance for when the policy gets creative.
More from coding & agent
- Ampersand Claims 92.2% on SWE-bench Pro HARD-51 at 70% Lower Cost — niloofar_mire · 2026-11-06
- ServiceNow Brazil Release Touts 100x Gains with Autonomous Engineer — azavery · 2026-11-06
- LiteParse Integrates 1.3B MiniCPM-OCR, Beating Gemini Flash on ParseBench — llama_index · 2026-11-06
- Finix MCP Server lets AI agents create buyers, payment links, and search docs — modelcontextprotocol · 2026-11-06
- Agent Substrate ships an MCP server for fs/bash on actors with auto suspend-resume — rakyll · 2026-11-06
- Personal agents weekly: Tomo adds Shop Pay via MCP, Fo's Twilio calls span 100+ countries — heyneighbor · 2026-11-06