AI news for Thursday, September 3, 2026

OpenAI has connected ChatGPT Health to Epic, the electronic-health-record system used for more than 325 million patients. Clinicians can pull notes, lab results, medications and specialist documentation into a chat without leaving the chart. The access is read-only—ChatGPT cannot write back to the record—and a separate plug-in searches public sources including ClinicalTrials.gov, PubMed and DailyMed. The convenience is obvious; the trust boundary is the story.

OpenAI says Astra is its first model to cross the Preparedness Framework's critical cybersecurity threshold. It scored a perfect result on ExploitBench and autonomously found and exploited two zero-days in modified tests. OpenAI plans a limited release, gating the strongest cyber capabilities to selected partners while adding chain-of-thought monitoring, jailbreak detection and containment-escape evaluations. The model's offensive capability—and who gets access—is the story.

South Korea's sovereign-AI program targets 8.4GW of data-center capacity by 2029 and 18.4GW by 2035, with planned investment totaling $919 billion, according to SemiAnalysis. The first phase allocates capacity across SK Group, GS Group and Naver while a government tournament chooses a national foundation-model champion. This is not another single-campus announcement; it is industrial policy spanning power, compute and domestic models.

A Texas sheriff's report about searching more than 80,000 Flock cameras for a woman who self-administered an abortion contains its own disclosure: the report was generated with Axon's Draft One. The product turns body-camera audio into police reports after an incident. Here, two automated systems sit in the same chain—one locating a person, another writing the official record. That combination deserves more scrutiny than either tool in isolation.

Anthropic's Enterprise Frontier Safeguards combines zero-data-retention model use with misuse monitoring whose activity logs stay in a customer's S3, Azure Blob or Google Cloud bucket under customer-controlled keys. Anthropic can evaluate risk without taking custody of those logs. For regulated buyers, this is the missing middle between having no monitoring and handing sensitive telemetry back to the model vendor.

CrowdStrike's SafeMind continuously runs Red Tempest, an offensive model trained partly on 15 years of incident-response data, against Blue Solano, which remediates the attack paths it finds inside a digital twin of a customer's environment. The loop continues until no viable path remains. The pitch is continuous red teaming at machine speed; the open question is how faithfully the twin represents the live environment.

Uber is lobbying beside driver unions for rules that keep humans in the network as autonomous fleets expand—the mirror image of its decade-long fight against organized labor. The Financial Times reports proposals including an 85% human-driver quota in New Jersey and a hybrid-network rule in Washington, DC. Uber's own estimate says one autonomous vehicle can displace roughly four drivers. The alliance is tactical, but the pressure it reveals is real.

An independent researcher says a three-week scrape of TikTok's private Android API produced 4.5 billion video records, now posted as a 289GB dataset on Hugging Face. The uploader acknowledges the collection violated TikTok's terms and prohibits identity, profiling or targeting uses in the dataset card. Those restrictions are policy, not a technical barrier—which is exactly why this matters.

Alibaba has refreshed Qwen3.8-Max without changing its 2.4-trillion-parameter size, 1M-token context window or API price. The company says the 0902 checkpoint is better at coding and Cowork-style office tasks and lifted CodeArena by 22 points to 1,691. Treat the leaderboard result as vendor-reported; the practical signal is a post-training gain rather than a larger model.

The U.S. Justice Department has filed a statement of interest backing OpenAI and Microsoft's argument that training models on copyrighted text can qualify as fair use in The New York Times lawsuit. The brief does not decide the case, but it arrives days before summary-judgment motions and puts the federal government behind the industry's central legal defense. The court still decides; publishers and model makers now have a clearer view of Washington's position.

Meta is backing away from using AI-adoption dashboards and token consumption in performance reviews after employees challenged the practice in court. At the same time it is pushing Hatch, a new internal agent, across the company. The contradiction is the point: mandatory measurement is retreating, but the pressure to produce with AI is not.

Kirkland & Ellis has committed $500 million to build custom AI systems with Palantir instead of relying only on off-the-shelf legal tools, the Financial Times reports. More than $100 million is planned for year one. The first product automates private-equity fund formation—documents, side letters, investor terms and compliance. This is what AI spend looks like when a professional-services incumbent decides workflow ownership is strategic.

Anker's Eufy MindBase moves home-camera analysis off the cloud: a 26-TOPS chip, 64GB of internal flash and support for up to 48TB of external storage run its TrueSmart Agent locally. It launches with new cameras and a security bundle. The useful bit is the architecture, not the 'AI brain' branding—local inference can reduce latency and how much private footage needs to leave the house.

MrBeast's next wilderness video will use Gemini to identify hazards and plan around weather, under a multiyear deal that also puts Fitbit and Google Health into future content. The world's largest YouTube creator is turning model demos into entertainment rather than placing ads beside entertainment. That is a much bigger distribution bet than another celebrity endorsement.