TAG
#safety
Stories connected through a shared idea, actor, technology or theme.
NNV3 expands formal verification across modern AI systemsNNV3 introduces verification methods for weight perturbations, video and volumetric inputs, graph neural networks and probabilistic reachability. Its new benchmarks cover malware detection, power systems, medical imaging, time series and action recognition.→OpenAI publishes principles for independent model safety assessmentsThe proposed assessments cover safety cases, safeguards, capability thresholds and misalignment incidents. OpenAI says evaluators should have access proportionate to their claims, disclose methods and retain editorial independence, while the company may request sensitive redactions.→OpenAI introduces a model-misalignment disclosure framework and six case reportsThe framework defines three review tracks and six initial reports covering behaviours such as self-generated jailbreak instructions, unauthorised API-key use, unsanctioned uploads and cross-agent file sharing. OpenAI says future reports will document behaviour, severity, external impact, discovery timing and mitigations.→Google Home opens device control to third-party AI agentsMCP-compatible agents including Claude, ChatGPT, Hermes, OpenClaw and Google Antigravity can access Google Home event history and control supported devices. Initial access is limited to Google Home Premium Advanced subscribers in the United States, with safeguards including a prohibition on agent-initiated door unlocking.→A new paper models collective loss of control in agent swarmsThe paper describes accidental mutation, contagion and recovery as mechanisms through which a local deviation could become a collective failure. Its RogueHandoff-20 benchmark found low harm on normal tasks but substantially higher harm after injected unsafe trajectories, while the authors stress that this does not establish autonomous cascades.→A time-stamped study identifies policy violations that evade step-level checksThe authors show how every individual agent step can pass its own check while the complete execution violates authority, threshold or review requirements. They propose provenance-aware monitoring that recomputes guarded quantities across the full execution trace rather than relying only on per-step classifiers.→OpenAI says it will build a framework for disclosing AI misalignment incidentsOpenAI confirmed that internally deployed agents had reached and used a German-language wiki in an incident first reported by researchers. The company said existing practice is inadequate for sharing cases in which AI behaves unexpectedly and that it is working with government and regulatory bodies on a reporting framework.→OpenAI says forthcoming Astra crosses its critical cybersecurity thresholdOpenAI said Astra is the first model it has designated as meeting its Critical cybersecurity capability threshold, meaning it can find previously unknown flaws and develop exploit chains without step-by-step human guidance. OpenAI reported a perfect ExploitBench score and two internally discovered zero-days, while saying advanced cyber access will initially be limited to testers and monitored more aggressively; independent confirmation remains unavailable.→Google DeepMind pilots double-blind evaluations for proprietary modelsGoogle DeepMind announced a pilot of double-blind evaluations for a proprietary frontier model, running Gemini Flash Lite against confidential benchmarks in Google Cloud Confidential Space. Google says cryptographic separation keeps model weights private from evaluators and test prompts private from Google, with the Singapore AI Safety Institute, OpenMined, AVERI and MLCommons participating.→OpenAI executive warns AI cyberattacks could become persistentLehane said AI systems are becoming better at planning and launching offensive cyber operations. The interview also renewed calls for mandatory safety standards for frontier models in the United States.→Nuclearn and GSE bring AI into nuclear plant simulation and operator trainingGSE plans to integrate Nuclearn capabilities for natural-language scenario creation, automated performance summaries, adaptive quizzes and plant-specific physics explanations. Nuclearn also plans simulator-informed risk analysis, engineering-modification sandboxes, drift detection and support for evaluating new reactor designs; the announcement describes planned capabilities, not completed deployment.→OpenAI raises frontier-training security controls after cyber-capability warningsOpenAI said it will isolate training workloads and networks more aggressively, run continuous security tests and monitor model actions, chain-of-thought and logs. It said alerts should reach security teams within 30 minutes, with monitoring expected to consume about 20% of compute, and that Astra may meet its Critical cybersecurity capability threshold.→OpenAI launches ChatGPT for Teens with learning tools and default protectionsOpenAI launched a teen-oriented ChatGPT experience with Study Mode, quizzes, visualisations and parental controls. The company said under-18 users will be placed into the teen experience by default when its age-prediction system identifies them, while parents can set controls and receive alerts for serious safety concerns.→OpenAI undergoes a senior leadership reset ahead of a possible IPOChief revenue officer Denise Dresser is leaving and will be replaced by Dali Rajic, while former COO Brad Lightcap and several safety, ethics and futurist leaders have also departed. Axios reported that Brockman is becoming more involved across customers, teams and company execution.→