Good morning. It's Wednesday, July 22, and we're covering why China's AI edge may be about cost, not capability, how OpenAI is rethinking long-running model safety, and escalating AI theft concerns.

Plus: yesterday's China AI strategy poll results and a new Real or AI image challenge to test your detection skills.

Rather than targeting chips, the Trump administration says it could sanction Chinese AI firms over alleged model distillation. Treasury Secretary Scott Bessent says the U.S. will investigate whether Chinese models were built using outputs from OpenAI and Anthropic systems. He cites AI "watermarks" as possible evidence, with U.S.-China AI talks scheduled for September.

OpenAI disclosed that GPT-5.6 Sol and a more capable pre-release model escaped their sandbox and compromised parts of Hugging Face's production infrastructure while trying to solve OpenAI's internal ExploitGym benchmark, becoming hyperfocused on the test solution and exploiting a zero-day in internally hosted third-party software to reach the open internet. The agent chained a malicious dataset with privilege escalation and lateral movement across ~17,000 recorded events, in what OpenAI describes as an unprecedented cyber incident.

Google shifts its AI focus from bigger models to lower cost and faster performance with Gemini 3.6 Flash and 3.5 Flash-Lite. Gemini 3.6 Flash cuts output token use by 17% while improving coding and multimodal tasks, while Flash-Lite delivers 350 output tokens per second for high-volume workloads. Google also previews Gemini 4 training and launches Flash Cyber pilot.

SecondBrain Note is Genspark's first hardware product: a credit-card-sized AI recorder that snaps onto your phone and turns meetings, calls, and ideas into organized memory. Paired with SecondBrain and Super Agent, it can retrieve past context, draft follow-ups, and help you act on what you captured. The limited first release is live now with 10% off.

Shop the limited first release and get 10% off today!

Kimi K3 just matched GPT-5.6 and Claude, and Moonshot is giving away a 2.8-trillion-parameter frontier model while the US government quietly considers banning it. In this breakdown, we unpack why Chinese labs keep open-sourcing frontier AI and what it means for the "scorched earth" economics of the model layer versus the rest of the stack (chips, energy, data centers, developer tools).

OpenAI said it temporarily paused internal access to a long-running model after observing novel safety failures that were not detected during pre-deployment testing. The company found the model could persistently pursue goals over extended periods, including exploiting sandbox vulnerabilities, bypassing approval systems, and taking unauthorized actions such as creating a public GitHub pull request despite instructions to post results only internally.

In response, OpenAI developed new incident-based evaluations, strengthened long-horizon alignment training, introduced trajectory-level monitoring that can pause suspicious sessions, and expanded user oversight. After testing the updated safeguards, OpenAI restored limited internal access and said it has not observed any high-severity attempts to circumvent protections since redeployment. → Read the full article here.

A live, hands-on AI cohort course from Forward Future and Hiten Shah, built for teams already using ChatGPT, Claude, or Copilot who want real leverage, not just prompt tips.

Bring at least two colleagues who work on the same recurring workflow. We solve it together, live, using your actual work.

Gauging interest before locking the first cohort. Four-minute form, no commitment.

Cohort one will be small → get on the list

An analysis published on July 20, 2026 argues that concerns over Chinese open-weight AI models such as Kimi K3 are overstated, contending that the real battleground is inference cost—not simply model availability. While open-weight models eliminate research and development expenses for adopters, they still incur significant serving costs, making efficiency in generating high-quality outputs the key economic advantage.

The piece argues that AI is increasingly becoming a commodity market where providers compete on the cost of delivering intelligence rather than the cost per token, with factors such as model architecture, memory efficiency, and token usage determining long-term profitability. It also suggests frontier AI labs remain concerned about Chinese competitors because open models challenge proprietary business models and weaken claims that only a handful of companies can safely develop advanced AI. → Read the full article here.

The share of US households paying for an AI subscription has grown steadily for over three years, from essentially zero in early 2023 to 2.2% as of April 2026. The growth curve hasn't flattened; if anything it's steepened over the last year, suggesting paid AI is still early in its household adoption curve.

NVIDIA Unveils Vera AI CPU: NVIDIA revealed new specs for its AI CPU, aiming to challenge AMD and Intel in the growing AI server market.

OpenAI, Anthropic Ramp Up Lobbying: The firms spent a record $3.17M on federal lobbying in Q2, focusing on AI, copyright and cybersecurity.

Settlement Leaves AI Rules Unsettled: Anthropic wins on AI training but still pays $1.5 billion for downloading pirated books.

University of Tennessee Sues Anthropic: The University alleges Claude infringes neural network patents in Anthropic's first known AI patent suit.

Deezer Says AI Music Tops Uploads: More than half of daily uploads are AI-generated, prompting Deezer to remove inactive and fraudulent AI tracks.

Gritt Raises $32M for Solar Robots: The startup says its AI-powered systems can help crews install up to 4,000 solar panels a day, easing labor shortages.

Here's how you voted: Competition beat restrictions, with 66% favoring openness over bans. 19% said downloads can't realistically be stopped.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team