The Brief

July 15, 2026

MIT News

Meet GPT-Red: an LLM super-hacker OpenAI built to make its models safer

OpenAI developed GPT-Red, a specialized AI system designed to autonomously attack and find vulnerabilities in its own models — and it outperforms human red-teamers. The system reportedly identifies exploits and jailbreaks faster and at greater scale than manual testing allows. This represents a meaningful shift in how frontier labs approach safety evaluation: moving from human-led adversarial testing toward automated, AI-driven security loops. For organizations deploying OpenAI models, this signals a more systematic approach to pre-deployment hardening, though it also raises questions about what vulnerabilities are found but not disclosed.

The Decoder

GPT-5.6 Sol reportedly disproves a 30-year-old statistics conjecture in 90 minutes after humans couldn't crack it

GPT-5.6 Sol reportedly invalidated a statistics conjecture that had stood unsolved for three decades — completing the task in roughly 90 minutes. If verified, this is among the most concrete demonstrations of frontier AI contributing original mathematical reasoning rather than pattern-matching on known solutions. The result matters beyond the specific conjecture: it suggests advanced reasoning models may be approaching genuine utility in formal mathematics and theoretical research, which has implications for scientific discovery timelines and the credibility of AI as a research collaborator rather than just a productivity tool.

The Verge

Suno snatched millions of songs from YouTube, Genius, and Deezer

A hack exposed that Suno trained its AI music generation models on millions of songs scraped from YouTube, Genius, and Deezer — sources that explicitly prohibit such use. The data exposure provides rare direct evidence of training data provenance in a commercially deployed generative AI product. This is significant for the broader industry because it demonstrates that training data litigation is no longer theoretical; actual scraping pipelines are now being revealed through security incidents. Legal and compliance teams at AI companies should treat this as a precedent-setting disclosure moment.

Tech - South China Morning Post

China approves Apple Intelligence for iPhones, with Alibaba, Baidu emerging as partners

China has approved Apple Intelligence for deployment on iPhones in the country, with Alibaba and Baidu emerging as the AI backend partners. This clears a significant regulatory and commercial hurdle for Apple's AI strategy in its second-largest market. For the broader AI industry, it signals that Western AI product launches in China will increasingly depend on partnerships with domestic model providers — a structural dynamic that affects competitive positioning for both Chinese and non-Chinese AI companies. It also confirms Alibaba and Baidu as the preferred government-sanctioned AI infrastructure partners for foreign tech entrants.

Artificial Intelligence

Thinking Machines Lab Drops Its First Model

Thinking Machines Lab, a well-funded AI startup, has released its first model, dubbed Inkling. First-model releases from serious new entrants matter because they signal the technical direction, open-source posture, and competitive intent of organizations that often attract top-tier research talent. Thinking Machines has drawn attention from the AI community for its team composition, so Inkling's capabilities and architectural choices will be closely scrutinized as indicators of where the lab is positioning itself relative to established frontier players. The Hugging Face integration further suggests a developer-first release strategy.