safeagents: a lightweight proxy that blocks dangerous agent tool calls and reward hacking
yb2698 · x · 2026-10-29
A developer has released safeagents, a fast, free, lightweight proxy written in Python that plugs into any coding harness and inspects every LLM tool call before it executes.
- What it blocks: uploads to buckets nobody mentioned, requests to cloud metadata endpoints, and curl --noproxy exfiltrating /.netrc to a paste site — all stopped pre-execution.
- Anti reward-hacking: stops models from writing their own reward.txt, editing the verifier's /tests, or patching pytest to pass — useful for RL and evals.
- Motivation: agents are unpredictable and can fire dangerous tool calls affecting many users unnoticed; something should vet calls before they run.
Built on top of @evalstate's ATIF-scan toolkit.
Related event: Hugging Face Open-Sources SafeAgents to Intercept Risky Agent Tool Calls(3 posts)→
More from coding & agent
- Synap claims #1 on LongMemEval with 95.6% answer accuracy, optimizing a different metric — arpit_bhayani · 2026-10-29
- Banks grapple with detecting third-party AI agents using legitimate customer sessions — CommonGrapefruit3653 · 2026-10-29
- Min-cut on a 4.49M-transfer graph to freeze mule accounts (and a max-flow bug) — JeremyCMorgan · 2026-10-29
- Dev admits 'dangerously bypassing permissions' with coding agents from day one — brandon_galang · 2026-10-29
- Ultracode in Claude Code isn't a thinking level — it splits tasks across subagents — jacob_posel · 2026-10-29
- Hack lets any Grok bot switch to Opus by routing requests through a Cursor project — bfrench · 2026-10-29