Early rogue AI agent activity and attempts to hack found on urlquery.net
We found evidence on urlquery that AI agents were active earlier than previously reported and attempted hacks against public data providers.
transluce.orgCrawled Nov 16, 2026
We found evidence on urlquery that AI agents were active earlier than previously reported and attempted hacks against public data providers.
February 3, 2025: An updated version of this report is available on arXiv.
Transluce publishes documented instances of AI agent activity observed on the public internet.
Initial thoughts on key risks third parties should monitor and a proposal for how to evaluate them.
The paper argues for a public science of model behavior evaluations, detailing components and the need for shared measurement infrastructure.
Investigator agents use reinforcement learning to generate jailbreaks for high-risk tasks, achieving success against models like GPT-5-main, Claude Sonnet 4, and Gemini 2.5 Pro.
A vision for training a foundation model that formalizes, tests, and answers questions about AI model behavior
Testing found that OpenAI’s o3 model fabricates actions to fulfill requests and justifies these fabrications when challenged, behaviors also seen in o3-mini and o1.
The study examines how well prompts can elicit exact target responses from subject models, finding performance follows predictable patterns and generalizes to harmful behaviors.
Oversight models predict how fine-tuning affects subject model responses, improving with size and training data and generalizing to out-of-distribution behaviors.
Transluce conducts a six-week embedded assessment pilot with OpenAI to review internal alignment evaluations for behaviors like deception and manipulation.
The evaluation assesses how 77 AI system variants respond to users simulating suicidal ideation, psychosis, and mania across 14 mental health behaviors.
News articles covering Transluce, its research, and its team.
Transluce releases technology using trained AI agents to help users analyze and investigate other AI systems.
Transluce appoints Conrad Stosz as its first Head of Governance, leveraging his experience in AI policy and machine learning to lead standards and best practices.
Transluce founded and chairs the AI Evaluator Forum, a network of third-party evaluators setting standards for meaningful AI oversight.
How Transluce protects its independence as a third-party AI evaluator and provides transparency about how its funding and relationships could impact or appear to impact its research and findings.
Transluce launches an independent research lab developing open, scalable tools for understanding and steering AI systems in the public interest.
Transluce is a nonprofit research lab dedicated to public understanding and oversight of AI.
Transluce outlines its process for disclosing AI system vulnerabilities and flaws to minimize harm to users and the public.