AI Today: A Model Attacks Riemann

#ai #news #dailydigest

Two unreleased frontier models were in the news this week for opposite reasons: OpenAI paused one because it got too good at hacking, and Anthropic pointed one at a 150-year-old math problem and let it run for 31 hours. The gap between "capability we're proud of" and "capability we're nervous about" is now measured in what the model happens to be aimed at.

Frontier Capability

An unreleased Anthropic model made progress on the Riemann hypothesis
It did not solve it — it significantly raised the lower bound for which the hypothesis is known to hold. The run is the interesting part: an Anthropic staffer with no serious math background prompted it, and the model tested 650 ideas across 60 subagents over 31+ hours, burning 31 million output tokens. Two of those 60 subagents produced the key mathematical ideas; the rest did development, validation, and documentation. Two in-house mathematicians confirmed the result, and it was formalized in Lean, which is the detail that makes this checkable rather than a press release.

OpenAI launched GPT-5.6-Cyber for vetted defenders
Built on GPT-5.6 Sol and trained for zero-day discovery and exploit chain development, it ships through a new Daybreak Red tier with system-level cybersecurity guardrails removed for approved researchers. It answers 95% of security prompts on OpenAI's own completion-rate benchmark, versus far lower rates for standard models, and OpenAI says it already found two unknown V8 bugs that could be chained to escape Chrome's heap sandbox, disclosed to Google. On Saturday I covered OpenAI halting work on Astra for crossing a "critical cybersecurity threshold." Three days later it shipped offensive capability to a vetted list. Both moves are defensible; together they're a bet that gatekeeping the access list is the control that matters.

Provenance

Anthropic will watermark text from its models
The watermark is embedded at the model level rather than added afterward, so it survives copy-paste and "may persist through some editing" — how much editing kills it is unanswered. Every Claude model released after August 2 carries it automatically, across the API, Claude, Claude Code, Cowork, and Claude Tag, with older models to follow; files use the C2PA standard. The August 2 start date is not a coincidence: that's when the EU AI Act's transparency code took effect.

Business & Industry

Nvidia lined up over $500B in third-party financing
Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR are building dedicated capital pools so hyperscalers, labs, and enterprises can buy Nvidia hardware without draining their own balance sheets. Jensen Huang told CNBC he approached exactly those six firms and none said no, and described his chips as an "investable asset." Read that as compute risk migrating from Nvidia's customers onto Wall Street's books.

General Catalyst led a $1.1B round into two-month-old River AI
xAI co-founder Igor Babuschkin's startup came out of stealth in June. AMP PBC co-led, with Nvidia, AMD Ventures, Y Combinator, and Temasek in. The product is an API for reinforcement learning and fine-tuning on open models — personally trainable assistants rather than worker replacements. Babuschkin's framing: agents as "guardian angels: quietly present, on your side."

Brad Lightcap is leaving OpenAI
CFO from 2018, COO from 2022, moved to special projects earlier this year, now out to "start something new." No named successor. He follows Fidji Simo in July, plus Bill Peebles and Kevin Weil — four senior departures while the company preps an IPO.

Gemini passed 1 billion monthly users
Sundar Pichai announced it Tuesday; Google's Q2 call had reported 950 million. Daily actives tripled year over year, 63% of users touch voice, 100M+ are on iOS, and the app generates 150 million images a day. ChatGPT crossed the same line in June.

One item I dropped: several roundups are running the Anthropic–Google–Broadcom 3.5GW TPU deal as August news. It was announced in April.