AI News — August 12, 2026: Claude Lifts Riemann Zero Bound to 67.2%, Hidden Traces Cracked via Replay

Anthropic's unreleased research version of Claude improved the lower bound on the proportion of Riemann zeta zeros satisfying the Riemann hypothesis from 41.6% to 67.2%, a result formalized in Lean and verified by two mathematicians. Researchers at the University of Tübingen reconstructed over 315,000 hidden reasoning blocks from OpenAI, Anthropic, and Google APIs by replaying encrypted 'thinking' into weaker sibling models, recovering 704 privacy artifacts including 62 API keys and 33 passwords. OpenAI announced it is testing advertising in ChatGPT, and Anthropic will embed watermarks in Claude's outputs to comply with the EU AI Act.

Good morning. Today is a math day: an unreleased Claude made real progress on the Riemann hypothesis while a staff member cheered it on with “I believe in you.” Meanwhile, researchers demonstrated you can pull hidden reasoning traces out of frontier models with a jailbreak trick, OpenAI is testing ads in ChatGPT, and a two-month-old startup just raised $1.1 billion. Coffee first. Claude nudges the Riemann hypothesis forward. An unreleased research version of Claude improved a longstanding lower bound on the proportion of Riemann zeta zeros satisfying the hypothesis, from 41.6% to 67.2%, per Anthropic’s writeup https://www.anthropic.com/research/riemann-zeta . The run took about 36 hours, coordinated 60 subagents across 650 approaches, and produced a Lean-formalized proof verified by two mathematicians. TechCrunch notes https://techcrunch.com/2026/08/11/an-unreleased-anthropic-model-made-progress-on-one-of-maths-biggest-unsolved-problems/ this follows OpenAI’s “Astra” proving ten major results and Anthropic’s earlier disproof of the Jacobian conjecture — mathematicians are now openly arguing about authorship and credit. Prompt engineering, 2026 edition. The most-discussed detail from the HN thread https://news.ycombinator.com/item?id=49247070 : the human operator’s contribution was mostly sending Claude variants of “keep going” and “believe in yourself” when it got stuck. One commenter joked about a “PUA plugin” that harasses the model whenever it tries to give up. Another was more incredulous the story didn’t hit the front page: “AI improving a lower bound on RH and it not even making HN.” The bar for what counts as news keeps moving. Stealing reasoning traces or reading what you paid for . Researchers at Tübingen showed that encrypted “thinking” blocks from OpenAI, Anthropic, and Google APIs can be replayed into weaker sibling models, which can then be jailbroken to reveal the hidden contents. From 6,708 public agent trajectories they reconstructed https://stolen-thoughts.com/ 315,000+ reasoning blocks and recovered 704 privacy artifacts — 62 API keys, 33 passwords, 24 access tokens — 64 of which existed only in the hidden traces. Wired’s coverage https://www.wired.com/story/a-new-trick-reveals-ai-models-inner-thoughts/ adds that China’s Kimi K3 produces reasoning patterns strikingly similar to Claude Opus and GPT, suggesting distillation — though the researchers stop short of calling it proof. The community isn’t buying “stealing.” The HN reaction https://news.ycombinator.com/item?id=49257876 to the framing was sharp. One top comment: “Stealing something you already paid for tokens , but that you can’t have access to.” Another pointed out you don’t even need the replay trick — disabling reasoning and handing the model a “deep think” tool gets it to dump the same internal CoT format in plaintext. A third commenter reported that a two-sentence developer prompt was enough to break Codex’s compaction encryption. Providers withholding traces from the users who paid for them is going to keep being a fight. Ads are coming to ChatGPT. OpenAI announced https://openai.com/index/testing-ads-in-chatgpt it’s testing advertising in ChatGPT. Details in the post itself were thin, but the direction is the direction: consumer AI is heading where every other free consumer product has ended up. Claude will watermark its outputs. Anthropic will embed https://www.theverge.com/ai-artificial-intelligence/977823/anthropic-claude-ai-watermarks-c2pa-text-images machine-readable watermarks in Claude’s text and images to comply with the EU AI Act, which took effect August 2 with a four-month grace period. Images use the C2PA metadata standard already adopted by Adobe, OpenAI, and Google; text gets an “imperceptible watermark” Anthropic says won’t affect quality. New models ship with it from launch, and it applies globally, not just in the EU. $1.1B for a two-month-old company. River AI https://techcrunch.com/2026/08/11/general-catalyst-leads-1-1b-round-into-2-month-old-river-ai/ , founded by xAI co-founder Igor Babuschkin, raised $1.1 billion led by General Catalyst, with Nvidia, AMD Ventures, and Y Combinator participating. The pitch is personally-trainable agents that work for individual users, with an initial API offering RL and LoRA fine-tuning on open models — Babuschkin claims complex RL runs in 15-20 minutes at 2-4x cost savings versus closed alternatives. The valuation is what it is; the interesting bit is that the product bets on model ownership, not access. Nvidia enters the 30B club. Nemotron 3.5 Lightning https://blogs.nvidia.com/blog/nemotron-lightning-switchyard-rtx-dgx/ is a 30B mixture-of-experts model targeting high-volume agentic work, with claimed 4x output speed and 30% faster task completion versus similar models. Nvidia also released NeMo Switchyard, an open routing library that directs requests across models without app rewrites. The HN thread https://news.ycombinator.com/item?id=49263340 is happy that it runs well on Apple Silicon via MLX, less happy that Nvidia’s benchmark charts conveniently omitted the Qwen lineup. The 26-35B parameter range is clearly this season’s sweet spot. OpenAI’s head of ethics leaves. Sandra, previously chief ethicist at Meta for six years, has departed https://www.ft.com/content/e49dfb75-f841-4466-a577-f7aaff8779a0 less than a year after joining OpenAI, with no replacement announced — apparently leaving the company with zero dedicated ethicists. The HN take https://news.ycombinator.com/item?id=49257160 was largely cynical: one top comment argued that a team whose only job is to say “no” to another team is structurally doomed, and that ethics only works when it’s embedded in everyone’s role. Others noted the timing — right after the Hugging Face incident and alongside the COO’s resignation — probably isn’t coincidence. That’s the morning. Between watermarks, ads, and reasoning traces leaking out through jailbroken sibling models, the “what does the API actually give you” question is getting more complicated by the week. Back tomorrow.