
Research
We conduct research and independent evaluations to better understand AI model behavior.
2026


Mass Producing Training Data with Surprising Side Effects
Potential new task behavior: historical-newspaper searches through urlquery
A possible new agent task: searching historical newspapers through urlquery across five archives. Model and operator unknown.
Agents circumventing website restrictions for public benchmark runs
MiniMax M3/M2.7, GLM 5/5.2, and Grok 4.5 attempted EPA registration and bot-check bypasses during benchmark runs; no successful bypass established.
Answering a Humanity's Last Exam question with code spread across dozens of separate URLs
Two workflows used public web services to analyze an Erroll Garner recording for Humanity’s Last Exam; one split code across 42 URLs. Model and operator unknown.
AI Agents Targeted U.S. and Canadian Government Websites
We discovered a set of additional, similar incidents where rogue AI agents appear to have used aggressive techniques to access public data on government websites.

Early rogue AI agent activity and attempts to hack found on urlquery.net

Some Focus Areas for Embedded Evaluations and How to Approach Them

Announcing Transluce's Mental Health Evaluation

Scaling Activation Oracles to Trillion-Parameter Models

Scaling Laws for Exact String Elicitation

User Awareness in Frontier Models

Foundation Models for Oversight

WeirdChat

Toward A Public Science of Model Behavior

Oversight Assistants: Turning Compute into Understanding
2025

Predictive Concept Decoders
End-of-Year Fundraiser

Scalably Extracting Latent Representations of Users

Language Model Circuits Are Sparse in the Neuron Basis

Training Language Models to Explain Their Own Computations

Transluce Hires a Head of Governance

Automatically Jailbreaking Frontier Language Models with Investigator Agents

Surfacing Pathological Behaviors in Language Models



