AI Briefing

Today's theme is authority arriving before the guardrails. Anthropic published a number nobody else has dared to: Claude now leads roughly a quarter of the work building Claude's successor. Meanwhile three researchers used Claude to walk into OpenAI's internal code repo, Google handed a Gemini agent its own verified Google account and the keys to your family calendar, and Brussels and a New York courtroom both started asking who pays for the last four years of scraping.

01ResearchAnthropic ↗

Anthropic Says Claude Now Leads 26% of the Work Building Its Own Successor

Anthropic published a prototype R&D Automation Index showing Claude 'leads' 26% of the company's AI research and development work as of August 2026, up from under 1% in February. 'Leads' means Claude completes most of a task end-to-end from a high-level prompt with a human supervising, and over 90% of the work now sits at 'AI collaborates' or above. It matters because no frontier lab has ever published this metric, and Anthropic is proposing it precisely so outsiders can watch how close the industry is getting to recursive self-improvement. The company also disclosed roughly 30,000 agents working internally at any one time, with only about 6% of AI R&D compute going to safety.

02IndustryTechCrunch ↗

Three Researchers Broke Into OpenAI Using Claude, and Opus 5 Cracked It Overnight

A three-person team at startup Hacktron AI chained a memory bug in the libheif image library with an SSO flaw to take over OpenAI employees' ChatGPT accounts, then reached an internal OpenAI GitHub repo and opened a harmless pull request as proof. The detail that should worry people: Claude Opus 4.8 failed across several sessions to build a working exploit, and within hours of Opus 5's release the same problem was solved. OpenAI fixed the flaws and paid a $6,500 bounty, but the episode shows a single model upgrade moving an attack from impossible to routine. As Hacktron's founder put it, work that once took months can now take days.

03PolicyTechCrunch ↗

Microsoft's Own Data Showed Copilot Cutting New York Times Clicks by 93%

Newly unredacted filings in the New York Times case against OpenAI and Microsoft show Microsoft's director of Applied Science calling AI scraping 'the largest theft of labor in human history' in a January 2023 memo, and describing a Copilot-driven 'doom loop' the following year. Microsoft's own measurements put the click-through drop to nytimes.com at up to 93% versus traditional Bing search. That number matters more than the quote: fair use turns partly on whether a product substitutes for the original, and OpenAI's head of ChatGPT reportedly called the products 'largely substitutive' in internal messages. Much of this comes from the Times' brief rather than the still-sealed exhibits, so context is incomplete.

04ProductGoogle ↗

Google Gave the Family Assistant Its Own Verified Google Account

Google Labs expanded CC from a personal daily-brief agent into a household one that up to six family members share, and the design choice worth noting is that CC now has its own verified Google Account with a distinct identity and permissions model. Members choose which senders to auto-forward, and CC sorts mail, updates a shared calendar, builds a 'Your Day Ahead' brief, and fills in permission slips and registration PDFs. Each instance runs in an isolated cloud environment on Google's Antigravity harness and current Gemini models. It is a bet that the next consumer AI interface is not a chatbot you open but a persistent account that acts on your behalf between the humans in your house.

05PolicyThe Next Web ↗

Brussels Wants AI Companions Switched Off by Default for Every Minor in Europe

The European Commission adopted a draft EU KIDS Act that bars under-13s from social media entirely, restricts 13- and 14-year-olds to parent-managed 'mini accounts' capped at an hour a day, and switches AI chatbots off by default for anyone under 18. Companion bots would also be barred from designs that create emotional dependency. The structural change is the burden of proof: large platforms would have to submit a safety plan to the Commission and an independent auditor before launching new features, rather than regulators proving harm after the fact. Fines reach 6% of global annual revenue. It is still only a proposal and needs Parliament and member-state agreement.