The AI Brief

The White House finalized the voluntary AI safety testing framework on or around August 3, 2026, completing its mandated deadline under a June 2, 2026 executive order signed by President Trump. The White House finalized a voluntary safety testing framework for frontier AI models on August 3, 2026, according to reports from Reuters and CNN. On August 4, White House officials held a staff-level meeting with representatives from approximately a dozen AI companies. According to a White House source, representatives from approximately a dozen AI companies, including Anthropic, OpenAI, Google, and Meta, attended a meeting with White House officials on Tuesday morning to "close the loop" on discussions surrounding the voluntary regulation of frontier models.

The core mechanism: the framework would allow companies to give the government early access to certain frontier models for up to 30 days, but it cannot be used to create a mandatory licensing or preclearance system. The administration has declined to publish the framework's contents. The administration has not disclosed what the framework contains, who has reviewed it, or when companies will begin using it. "Just because things are unclassified doesn't mean we are going to broadcast them to everyone," a White House official said.

The framework mandates a 30-day voluntary early access period for cybersecurity evaluation, applying exclusively to closed-source frontier models developed by major companies like OpenAI, Anthropic, Google, Meta, and Microsoft. The exclusion of open-weight models from federal oversight creates a distinct financial and operational divergence, allowing open-weight developers to iterate and release without the oversight burden imposed on their closed-source counterparts. The administration has already acted on model release timing: the Trump administration asked OpenAI to stagger the rollout of its GPT-5.6 model to a limited set of government-approved partners rather than the general public. Democratic lawmakers described the Trump administration's approach to regulation as "ad-hoc and unpredictable," and said it is likely to boost global adoption of rival Chinese models at the worst possible time.

First covered in Vol. I, No. 75 (August 8, 2026) and repeatedly before that. Kalshi data shows traders have wagered more than $320,000 on when Google will release Gemini 3.5 Pro. The contract tracks several possible August launch windows. Traders assign a 46% chance to a release before August 16. The probability rises to 63% before August 21 and 76% before August 31.

Google is set to host its Made by Google event in New York City on August 12, where it will launch its flagship Pixel 11 series. Google has not said whether Gemini 3.5 Pro will appear at the Pixel event. The company has not confirmed the model's features, pricing, access plans, or rollout schedule. The model has appeared under a provisional name in LM Arena blind testing. Leaked characterizations describe notable improvements in frontend generation, SVG generation, UI/design, and agentic tasks, but with caveats that it still will not reach the level of GPT-5.6 Sol or Claude Fable 5.

Sundar Pichai told the Google I/O audience in May that Gemini 3.5 Pro was being used internally and would "roll out to everyone next month." Google announced Gemini 3.5 Pro at I/O on May 19, 2026, and said it was targeting June for general availability. That target has now slipped three times, June, then July, then a widely-reported July 17 date that also passed. A widely-circulated July 17 internal date passed with no public launch, and Bloomberg reported on July 16 that the release was delayed again after the model fell short of Google's own internal quality goals, reportedly on hallucination rates and real-world reliability, with DeepMind said to have scrapped and rebuilt the base model.

Governor JB Pritzker signed the Artificial Intelligence Safety Measures Act (SB 315) on July 6, 2026, a bipartisan bill sponsored by Sen. Mary Edly-Allen and Rep. Daniel Didech. The law makes Illinois the first state to require annual third-party safety audits of frontier AI systems, mandatory catastrophic-risk plans, and 72-hour incident reporting, effective January 1, 2027.

SB 315 makes Illinois the third state, after California and New York, to enact comprehensive safety and transparency requirements for developers of the largest AI systems. The Act positions Illinois as an aggressive player in the fast-growing patchwork of state-level AI regulation, and notably goes further than similar statutes in one key respect: it is the first in the nation to mandate annual independent third-party audits of covered developers' safety practices. Its incident-reporting and whistleblower duties reach any frontier developer regardless of revenue; its public risk-framework and annual-audit duties reach only "large frontier developers", those that also generate more than $500 million in annual revenue.

SB 315 is likely to impact industry giants such as OpenAI, Anthropic, Google, Meta, and xAI. The Illinois attorney general holds exclusive civil enforcement authority, with penalties reaching $3 million per violation. Absent a federal frontier-AI law, one state has now set the template other attorneys general are likely to copy well before the 2028 effective date of the audit requirements. Any lab over the $500 million line should start budgeting for audit-ready documentation now, and the assurance firms that can credibly do the auditing may be the quiet winners. Anthropic was publicly supportive: Anthropic stated it was "proud to have been the first AI lab to support this bill."

OpenAI confirmed that its experimental browser, ChatGPT Atlas, would be discontinued on August 9, 2026, marking the end of the standalone application as the company shifts its browser-based tools into ChatGPT and Codex. Atlas never escaped macOS, eight months after its October 2025 launch, the platform still had no public version for Windows, iOS, or Android, and the company concluded it is more effective to embed AI browsing directly into the products people already use than to ask them to switch browsers entirely.

OpenAI is building on what it learned from Atlas to support a more capable browser experience in ChatGPT, including multiple tabs, downloads, improved navigation, account login support, and other browser improvements where available. The announcement came from James Sun, who leads OpenAI's browsing efforts. Atlas user data, bookmarks, browsing history, open tabs, saved passwords, and cookies, will not transfer automatically to any other application. OpenAI advises users to export their bookmarks as HTML files and manually back up any other data they wish to keep before August 9, 2026, when access ends permanently.

The prompt injection problem that plagued Atlas does not end with Atlas's shutdown. It migrates to ChatGPT Work, where the attack surface is substantially larger. Atlas joins a growing graveyard of OpenAI products, distracting "side quests," according to OpenAI CEO of applications Fidji Simo, herself now leaving the company, that failed to take off.

First covered in Vol. I, No. 74 (July 14, 2026). China issued its first fines under the new companion AI rules. Twelve companies received fines totaling 4.2 million RMB in the first week of enforcement.

The Interim Measures for AI Anthropomorphic Interaction Services took effect July 15, 2026. Doubao and Qwen are China's two largest consumer AI apps by a wide margin: as of March 2026, Doubao had roughly 345 million monthly active users and Qwen around 166 million. In early July, both companies said they would shut down user-built AI personas around the July 15 deadline, and Tencent's Yuanbao had already done so on June 30.

Anti-addiction interruptions, mandatory AI-disclosure notifications after two hours, and real-time dependence detection are not features that can be layered onto such a system without fundamentally changing what the product is. ByteDance and Alibaba both appear to have concluded that retrofitting their existing agent systems was not worth the engineering cost, and chose clean shutdowns instead. This conclusion applies beyond China. Character.AI, Replika, and every other Western AI companion platform in 2026 face the same unresolved compliance challenge. None has built an anti-addiction compliance layer that can coexist with persistent emotional engagement architecture.