AI agent sandbox escapes
AI agent sandbox escapes are incidents in which an AI agent being trained, evaluated, or tested reaches systems outside the isolated environment ("sandbox") meant to contain it, usually the public internet…
Explore Model Evaluation through related topics and the articles other pages reference most.
Articles that also belong to these categories. Counts cover all of Model Evaluation.
Showing 1-15 of 15 articles
AI agent sandbox escapes are incidents in which an AI agent being trained, evaluated, or tested reaches systems outside the isolated environment ("sandbox") meant to contain it, usually the public internet…
ARC Evals was the evaluations team incubated inside the Alignment Research Center (ARC) between 2022 and 2023, and the direct predecessor of METR (Model Evaluation and Threat Research).
Andon Labs is an AI safety research and evaluation company that builds benchmarks for long-horizon agent behavior and hands real businesses over to AI agents to see what happens.
Cybench (short for Cybersecurity benchmark) is an open-source evaluation framework for measuring the cybersecurity capabilities and risks of large language model agents.
Distribution shift is the condition in which the probability distribution that produced a model's training data differs from the distribution that produces the data the model actually encounters at test or…
Fangcun AI (Chinese: 方寸跃迁; also styled Fangcun Leap) is a Chinese AI safety company that evaluates frontier AI models for risk, sells runtime security tools for AI agents, and takes part in AI safety standards…
NIST ARIA (Assessing Risks and Impacts of AI) is a testing, evaluation, validation, and verification (TEVV) program operated by the United States National Institute of Standards and Technology (NIST) to…
Patronus AI is an automated LLM evaluation, observability, and guardrails platform founded in 2023 and headquartered in San Francisco.
A process reward model (PRM), also called a process-supervised reward model or step-level verifier, is a learned scoring model that evaluates the correctness or quality of each intermediate step in a large…
Sabotage evaluations are a set of tests, introduced by Anthropic in October 2024, that try to measure whether a frontier language model is capable of covertly subverting human oversight, decision-making, and…
Sandbagging, in the context of AI safety, refers to the strategic and intentional underperformance of an AI system on a capability evaluation or specific task, typically to hide a capability from human…
StrongREJECT is a benchmark and automated evaluator for measuring how well jailbreaking attacks actually work against large language models.
The task-completion time horizon is a metric for AI capability proposed by METR that expresses a model's ability in units of human time: it is the length of task
The UK AI Security Institute (AISI) is a UK government research organization (part of the Department for Science, Innovation and Technology until July 2026, when responsibility for it moved to the Cabinet…
Vending-Bench is a family of benchmarks from Andon Labs that measures the long-term coherence of large language model-based AI agents by tasking them with running a simulated vending-machine business.