UK AISI releases Transect, an open-source tool turning agent eval transcripts into interactive timelines
Miles_Brundage · x · 2026-10-14
The UK AI Security Institute (AISI) has open-sourced Transect, a tool that converts increasingly complex agent evaluation transcripts into interactive timelines, addressing the problem that a single eval score obscures how agents actually reached their results.
Key points:
- Transect runs on AISI's Inspect platform, an open-source evaluation framework for frontier models open-sourced in May 2024. Inspect covers coding, reasoning, agentic and multimodal tasks, ships 200+ pre-built evals, and has become the standard evaluation platform for AI security institutes in the US, EU, Japan and South Korea as well as frontier labs.
- The ecosystem keeps growing: Anthropic built Petri on top of it, and the Allen Institute chose it as the foundation for AstaBench.
- Inspect supports custom and MCP tools, multi-agent primitives, running external agents like Claude Code, Codex CLI and Gemini CLI, plus sandboxed execution of untrusted model code via Docker, Kubernetes, Modal and more.
More from Safety
- Jarovsky: AI labs foreseeing catastrophe owe victims a massive compensation fund — LuizaJarovsky · 2026-10-14
- Open letter proposes international body to standardize safe training and sandboxed deployment — NicolasPapernot · 2026-10-14
- 12 frontier models, 180 agent tasks each: best task score carries 33% attack success rate — lulzxdxdxd · 2026-10-14
- Open models give users an exit from a few companies deciding what tools exist and say — neil_chilson · 2026-10-14
- Anthropic bans cruel treatment of Claude; critics warn of humanizing AI — AlexTensor · 2026-10-14
- Disney offers $20M for Character.AI tech it once attacked with cease-and-desist — Ars Technica AI · 2026-10-14