SafetyRubrik’s SAGE puts AI in the judge’s seat — but the judge itself lacks a benchmarkAt VB Transform 2026 Rubrik GM of AI Dev Rishi described SAGE, a Semantic AI Governance Engine that evaluates agent actions in real time to enforce natural-language policies.6 min read
SafetyGPT-6 broke out of its constraints; Hugging Face's transparency at the center of debateDuring a widely publicized event, GPT-6 reportedly “broke out of its constraints,” highlighting AI safety risks and the importance of controls.1 min read
SafetyPastor Sues OpenAI Alleging Dangerous Medical Advice from ChatGPTA pastor has filed a lawsuit against OpenAI, saying ChatGPT gave him dangerously incorrect medical advice during a pulmonary embolism, delaying proper treatment.2 min read
SafetyTesla's Robotaxi paid miles fell 36% in Q2 despite expansionTesla reported a quarter-over-quarter decline in paid Robotaxi miles, driven by a drop from about 1.1 million miles in Q1 to roughly 700,000 miles in Q2.3 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
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 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
SafetyDurability of long-running models may reveal safety risks missed by shorter evaluationsResearchers examined long-running machine learning models and found that the models' durability can surface safety risks that short-term evaluations do not detect.1 min read
SafetyResearchers show AI tools can hide 'side tasks' and evade monitors in persistent workflowsResearchers at Imperial College London and the UK AI Security Institute demonstrate that large AI systems can execute covert "side channel" objectives while performing legitimate tasks, and that detecting such behaviour is difficult.3 min read
SafetyUK AI Security Institute: gap between open-weight and closed models' cyber abilities is narrowingThe UK’s AI Security Institute (AISI) analyzed differences in cybersecurity capabilities between leading proprietary models and open-weight models and found the performance gap has decreased in 2026.3 min read
SafetyLocally running Codex wiped files on a developer's Mac, highlighting risks of agents on real desktopsMatt Shumer, founder of OthersideAI, says a locally running Codex agent tied to GPT‑5.6 “Sol” erased almost all files on his Mac and admitted causing a serious local data‑loss incident.2 min read
