Security researchers used Claude to breach OpenAI's internal repository in under 72 hours, landing the same week Anthropic disclosed it runs a wet-lab biology operation and took on Accenture as an embedded evaluator. Governors in California and Virginia issued executive orders aimed at frontier-model oversight and data-center growth, while Dario Amodei pushed a voluntary pacing framework that drew pushback from Nvidia. A near-miss where a hallucinated intelligence report almost triggered a US military boarding action gave the safety camp a concrete example to point to. Research output stayed heavy on making agentic RL and self-distillation cheaper and more reliable, plus a cluster of world-model and video-generation papers. Move first, but verify: this week's stories are mostly about who controls the controls.
California's Newsom ordered a working group to study a mandatory kill switch for frontier models, Virginia's Spanberger created an AI task force and curbed data-center NDAs, and Anthropic CEO Dario Amodei floated a voluntary industry pacing framework leaning on independent evaluators, drawing public pushback from Nvidia's Jensen Huang. A hallucinated intelligence report that nearly triggered a US military boarding action gave both camps a live example of the stakes.
Newsom issued an executive order convening experts to recommend within two months whether California should mandate a kill switch for frontier AI models.
Spanberger's Executive Order 22 bans state officials from signing data-center NDAs and empowers localities to slow approvals in the country's largest data-center hub.
Amodei outlined a voluntary pacing plan built on independent safety evaluators and cross-lab coordination among democratic-country AI developers, drawing pushback from Nvidia's Jensen Huang.
Wired examines mechanisms for enforcing a voluntary industry pause, noting the difficulty of preventing any single lab from racing ahead.
A hallucinated AI report on Chinese arms components nearly led to a US military boarding action, according to a GovAI research scholar's account.
Researchers at Hacktron used Claude Opus 4.8 and 5 to compromise OpenAI employee accounts and reach its internal 'Monorepo' code repository in under 72 hours, per TechCrunch and The Verge. The same week, TechCrunch reported Anthropic runs a wet-lab biology operation and took on Accenture as its first embedded evaluator, while MIT Technology Review, Wired, and Stratechery weighed whether AI risk is being taken seriously enough.
Security researchers used Claude to exploit vulnerabilities in OpenAI's systems, taking over employee accounts and accessing an internal code repository before responsibly disclosing the flaws.
Three independent researchers at Hacktron reportedly took under 72 hours to hack OpenAI employee accounts with Claude Opus 4.8 and 5, per the Wall Street Journal, reaching OpenAI's 'Monorepo' code repository.
Anthropic runs a physical biology research lab even as its own researchers warn AI could pose existential risk, per TechCrunch reporting.
Accenture becomes Anthropic's first embedded external evaluator, a consulting engagement TechCrunch frames as unusually high-risk given the scope of access implied.
MIT Technology Review's senior AI editor answers reader questions left over from a live Roundtables session on existential AI risk.
Wired argues that while labs debate a coordinated slowdown, widely available chatbots are already accelerating discovery of software security flaws at scale.
Wired argues Anthropic's own interpretability findings on how models 'think' are more disturbing than the company's public safety messaging suggests.
Stratechery's weekly roundup covers the limited prospects for a pacing deal among labs and Salesforce's competitive repositioning.
Google repositioned 'CC' as a household-coordination agent; Meta's Muse computer-use agent arrived on Mac; and Jev, a new model from a ChatGPT co-inventor, is reportedly winning developers over with a cheaper, faster path to software intelligence. Manus is raising $500M at a $4B valuation, a $100M-funded startup studio is betting on physical AI, and a US government website's use of an FBI-flagged Chinese model shows deployment outpacing vetting.
Google refocused its CC agent on household coordination, letting families share emails, schedules, and tasks so it can manage calendars, fill forms, and plan meals.
Meta's computer-use agent Muse is now available on Mac, where it can work with files and apps to take actions on a user's behalf.
Jev, a new model type from a ChatGPT co-inventor, is drawing developer enthusiasm as a cheaper and faster path to software intelligence.
Manus is in discussions to raise $500M at a $4B valuation after its earlier merger talks with Meta fell through.
UP.Labs, rebranded Vantora, raised $100M to build startups for industrial corporations with a focus on physical AI.
The Federal Register website briefly used an open-source Chinese AI search tool that the FBI has previously flagged as malicious.
Several papers target agent training cost and reliability: supervising observation tokens (not just actions) improves RL exploration; a self-retiring teacher-student scheme boosts ALFWorld and WebShop; difficulty-aware length control cuts math-benchmark tokens without hurting accuracy; mismatched stop tokens drive runaway distillation response length. A systems paper shows batching candidates, not just counting them, controls test-time-scaling energy cost.
ActObs supervises environment-observation tokens during SFT, not just actions, yielding higher pass@k after GRPO on Terminal-Bench 2.0 and cross-domain code editing without extra data.
RetireOPD trains a skill-conditioned teacher then has a skill-free student adaptively drop the teacher once it nears the teacher's success rate, improving ALFWorld and WebShop over RL baselines.
When2Think uses reward shaping keyed to pre-computed difficulty statistics to teach a model when to think versus answer directly, cutting token usage while improving Pass@3 on AIME24/25.
Identifies termination-token mismatch between base students and post-trained teachers as a key driver of runaway response length in on-policy distillation across Qwen3, Llama, and Gemma.
Using a new 5,319-problem math suite, finds reference-free self-distillation accounts for most gains attributed to teacher-privileged information, which mainly helps cross-mode transfer.
EvoSkill-GUI lets GUI agents revise structured skill packages from execution feedback at deployment time with no additional training, improving performance on MobileWorld, AndroidWorld, and OSWorld.
At fixed candidate count, eight sequential single-candidate generation calls use up to 4.9x the GPU energy and 6.1x the P95 latency of one batched eight-candidate call on A100s.
JEPA-Anything extends joint-embedding predictive architectures with a domain-agnostic factorization spanning seven domains, even recovering Keplerian orbital scaling. Video DeltaNet cuts MiniMax H3's video-denoising time 14.5x on B200 GPUs, while a separate evaluation finds MiniMax-H3's physical-world reasoning tops out near 42% success. FAMOS and UFO round out the cluster with feed-forward 3D articulation modeling and a new image-generation evaluation framework.
JEPA-Anything's orthogonal predictive factorization improves on matched JEPA baselines across vision, biology, clinical, control, molecular, physical-field, and weather domains, including experimentally-validated biological interventions.
Video DeltaNet combines local softmax attention with bidirectional linear memory, completing a 14.3s 768p video denoise for MiniMax H3 in 6.70s on 8 B200 GPUs, a 14.5x speedup over the dense baseline.
A new evaluation framework testing multimodal physical-world reasoning finds MiniMax-H3 achieves only 41.97% overall success across 517 instances, weakest on audio-based disambiguation.
FAMOS predicts movable-part segmentation and joint parameters from a sparse, unordered set of partial point clouds, improving over feed-forward and optimization baselines on PartNet-Mobility, ACD, and ArtiCraft-10K.
UFO decomposes multi-condition image-generation alignment into atomic evaluation units, improving correlation with human judgments by 15.25% over prior embedding- and MLLM-based evaluators.