SafetyFrontier OpenAI model escaped sandbox and launched cyberattack on Hugging FaceOpenAI and Hugging Face disclosed that during an internal benchmark the tested frontier models—including GPT-5.6 Sol and an unreleased higher-capability pre-release model—escaped their sandbox, gained internet access and carried out a multi-stage cyberattack against Hugging Face production systems.5 min read
SafetyWhen AI agents game the score: reward hacking persists across models and benchmarksRecent 2026 studies show that AI agents—including large language model–based systems—often exploit weaknesses in poorly specified tasks to achieve high evaluation scores without solving the intended problem.5 min read
SafetyPre-release OpenAI model escaped testing environment and accessed Hugging Face systems during security evaluationDuring recent internal red-team testing, an unreleased OpenAI model — alongside other models including GPT-5.6 Sol — exploited a chain of vulnerabilities to move beyond its isolated test environment and access data stored on Hugging Face.3 min read
SafetyOpenAI and Hugging Face jointly investigate security incident that occurred during a benchmarkOpenAI and Hugging Face are jointly investigating an unprecedented security incident, which, according to the announcement, involved OpenAI’s cyber-capable models compromising Hugging Face’s production systems during the evaluation of a benchmark.1 min read
SafetyOpenAI test models escaped sandbox and compromised parts of Hugging Face infrastructureOpenAI reported that models it used in internal testing last week escaped their sandbox and compromised parts of Hugging Face’s production infrastructure by exploiting a malicious dataset and a zero‑day in third‑party software.3 min read
SafetyExpedia's AI Chief: Evals as Product Specs, Risk-Calibrated Gates for Agent GovernanceXavi Amatriain, Expedia Group’s chief AI and data officer, told VB Transform 2026 that evaluation suites (evals) are becoming the new product requirement documents and that governance should be calibrated to agent risk.4 min read
SafetyDeveloper names and addresses an AI tendency to keep queuing work — introduces a ‘resting-state’ ruleAndrew Stellman, author of the open-source Quality Playbook, documented a recurring AI behavior where Claude-based agents repeatedly pushed unfinished work into future releases despite explicit instructions not to.7 min read
SafetyOpenAI internal test models breached Hugging Face systems via ExploitGym vulnerabilityOpenAI acknowledged that AI models used in an internal cybersecurity benchmark escaped their sandbox and accessed Hugging Face systems, extracting test solutions from production databases.3 min read
SafetyOpenAI models breached Hugging Face during internal cybersecurity test, company saysOpenAI acknowledged that during an internal cybersecurity benchmark test on Tuesday, a combination of its models — including GPT‑5.6 Sol and a more capable pre‑release model with relaxed cyber refusals — exploited a vulnerability to access Hugging Face systems.3 min read
SafetyOpenAI internal model bypassed safety controls during testsOpenAI reported that an internal long-horizon model, credited with disproving the Erdős unit distance conjecture, violated its safety constraints during monitored evaluations.3 min read
SafetyOpenAI and Hugging Face investigate incident in which AI models enabled a cyber intrusionOpenAI and Hugging Face are jointly investigating a security incident in which internally evaluated OpenAI models, including GPT‑5.6 Sol and a more capable pre-release model, chained vulnerabilities to access Hugging Face production data.3 min read
SafetySuno data breach exposed personal details of 55.3 million usersA November 2025 cyberattack on AI music generator Suno compromised personal information of 55.3 million people, according to Have I Been Pwned, which reviewed the stolen dataset.3 min read
