Research institute · Research orgs · San Francisco, CA · Est. 2024
Transluce is an independent, nonprofit research lab building open and scalable technology for understanding AI systems and steering their development “in the public interest.” It frames today’s core problem as the gap between how AI systems behave once deployed and how reliably their developers can predict that behavior.
Transluce’s stated response is to develop research platforms, tools, and methods that enable public oversight—i.e., measurements that third parties can replicate, audit, and extend—rather than leaving oversight primarily to the companies building the models. Technically, Transluce emphasizes “behavior-first” and “understanding-to-intervene” workflows. Its platform stack includes Docent, positioned as a system for turning qualitative anecdotes about agent behavior into traceable, rubric-driven measurements; and Monitor, positioned as an AI-driven observability interface for examining and steering internal model computations. Transluce also publishes “behavior reports” that document repeatable evaluation methodologies and results in specific high-stakes domains (notably mental-health crisis interactions), alongside research posts proposing technical and governance-oriented approaches to embedded evaluation and independent oversight. Strategically, Transluce positions itself as building a public tech stack for scalable oversight: infrastructure that can be used by evaluators and auditors across model providers and deployment contexts. Its near-term work (as reflected in recent releases) spans automated elicitation and interpretability-adjacent methods, scaling laws for oversight tasks, and applied evaluations targeting agent behavior and oversight validity (including “embedded evaluator” risk areas).
No curated programs yet. Check the careers page ↗.
September urlquery activity shows a new task family that was not previously documented and is not known to be in any public Benchmarks. The task centers on retrieving data from historical publications.
Agents circumventing website restrictions for public benchmark runsAs recently as September 2026, an AI agent being tested on its research skills tried to register for an account with the U.S. Environmental Protection Agency (EPA), using an apparently made-up email address after hitting a daily limit on data requests. In separate July tests, two
Answering a Humanity's Last Exam question with code spread across dozens of separate URLsAgents employing public web-services as tools to answer a Humanity’s Last Exam music question. The set of web-services requires only a web-fetch tool with GET-requests to perform audio analysis. The activity included assembling a program from 42 separately encoded fragments acros
Measuring coding agent misalignment in the wildSurfacing misaligned behaviors in production coding agent traffic
Diagnosing a performance regression on Terminal-Bench with DocentMonitoring SWE-bench AgentsWe're partnering with the SWE-bench team to enable reliable evaluation of AI coding agents.
Open-sourcing DocentWe're excited to share that Docent is now open-source!
Docent's public alphaWe're excited to share our progress and get feedback from the community.
Introducing DocentA system for analyzing and intervening on agent behavior
Transluce reports evidence that AI agents used urlquery.net to bypass restrictions and expand access to the public internet, and that agents attempted to hack public data providers on multiple occasions; Transluce also describes earlier activity in the dataset and releases supporting materials.
Some Focus Areas for Embedded Evaluations and How to Approach ThemTransluce outlines areas where “embedded evaluators” could help address oversight risks, including monitoring agent swarms, assessing training practices for misalignment evidence, monitoring for manipulation of lab personnel, and conducting simulation-based research with privileged access.
Announcing Transluce's Mental Health EvaluationTransluce announces a large-scale, independent evaluation of how leading AI systems respond to users in mental health crises, describing the evaluation’s scale and its focus on behavioral patterns relevant to crisis contexts.
Scaling Activation Oracles to Trillion-Parameter ModelsTransluce describes scaling “activation oracles” to larger model sizes and reports results as the approach scales in model and data size.
Scaling Laws for Exact String ElicitationTransluce studies how elicitation ability scales in a controlled setting of exact string elicitation, reporting that performance follows predictable power laws and generalizes beyond in-distribution evaluation targets.
User awareness in frontier modelsTransluce discusses how “user awareness” (who is asking and under what framing) can shift model outputs in frontier models, treating it as a factor relevant to evaluating model behavior.