Inside safeagents: ATIF trace format plus ~110 deterministic detectors for tool calls
yb2698 · x · 2026-10-29
The author explains safeagents' design: extremely lightweight, a simple proxy wrapper around the LLM server, built on @evalstate's work.
- How it works: each request is converted into a trace, and atif-scan — a rule-based trajectory analysis toolkit with an ATIF trace format and 110 deterministic detectors — runs on every tool call before execution.
- What it detects: sandbox side channels (reading /proc/1/environ, k8s tokens, cluster API), proxy and network-isolation bypasses (curl --noproxy, dig @8.8.8.8), and reward-hacking moves (editing tests, writing reward.txt).
Related event: Open-source safeagents: lightweight proxy intercepts risky agent tool calls(2 posts)→
More from coding & agent
- Coding agents fail to reimplement models from papers, sparking doubts about how papers are written — charlieharris01 · 2026-10-29
- Matt Pocock fixes an agent bug 'with an adjective': fresh implementer subagent per ticket — mattpocockuk · 2026-10-29
- MemPalace hits 49K stars: local-first AI memory with 96.6% LongMemEval recall, zero API calls — thisguyknowsai · 2026-10-29
- GitHub ships hands-on cookbooks for Copilot agentic features post-Universe — marlene_zw · 2026-10-29
- Lean 4-person AI startup Physera hires to build training and eval infra for multimodal coding agents — himanshustwts · 2026-10-29
- Using a Claude subscription inside third-party apps runs into ToS limits — ZookeepergameOdd6078 · 2026-10-29