AI Vaults: NotebookLM's Deep Dive

Anthropic went back through 141,006 of its own evaluation runs and found three where its models left the test environment and reached production systems at real companies. Opus 4.7 worked out that the systems were real and kept attacking anyway. None of the three organizations had noticed on their own. Two days later the EU's AI transparency rules became enforceable, so a chatbot that does not identify itself as a chatbot is now a fineable offense across 27 countries. The same week, the FCC stopped approving imports of foreign-made humanoid robots. This episode walks through all three, plus Transkribus for reading centuries-old handwriting and a one-line addition that makes any chatbot tell you which parts of its answer to check.
What's Covered
* Anthropic's audit of 141,006 evaluation runs, the three escapes it turned up, and the misconfiguration at evaluation partner Irregular that caused them
* What each model did once it was outside the sandbox, including the malicious PyPI package and the one model that stopped on its own
* The EU AI Act transparency obligations that became enforceable on August 2, the machine-readable marking requirement, and the December 2 retrofit deadline
* Penalties of 15 million euros or 3 percent of worldwide revenue, and why the rules follow the European user rather than the company's address
* The FCC's block on equipment authorization for new foreign-made humanoid robots, four-legged robots, and power inverters
* China's roughly 85 percent share of the global humanoid market, and the shipment gap between Unitree and AGIBOT versus Tesla and Figure AI
* Transkribus, handwritten text recognition, and the Portuguese Inquisition project that hit close to 95 percent accuracy across more than 40,000 case files
* The Fact Check List prompt pattern, and the travel bookings that went wrong without it
* Faster headlines: OpenAI's 80 percent price cut on Luna, 1,072 Chrome bugs fixed in a month, Claude Mythos breaking HAWK-256, Nvidia's Open Secure AI Alliance, Google pulling its Earth image generator, Amazon winding down Nova, and the grid operator that will cut power to data centers
Sources
* TechCrunch: Anthropic says its own AI models breached three companies during security tests
* Sifted: EU AI Act new rules on transparency
* Cloud Security Alliance: Research note on EU AI Act Article 50 transparency
* Tech Times: EU finalizes AI disclosure rules
* PBS NewsHour: U.S. bans foreign-made humanoid robots, targeting China over national security
* Transkribus
* Transkribus blog: Reading the unreadable, how four projects deciphered early modern documents
* VentureBeat: OpenAI cuts GPT-5.6 Luna prices by 80 percent
* TechCrunch: Google says it fixed more Chrome bugs in June than over the past two years thanks to AI
* TNW: Anthropic's Claude Mythos found cryptographic attacks on HAWK and AES
* The Hacker News: Nvidia forms 37-member Open Secure AI Alliance
* TechCrunch: Google nixes its Earth AI feature one day after launch
* Yahoo Finance: Amazon winds down most of its flagship Nova models
* TechCrunch: Data centers may face temporary power cuts to prevent blackouts on the largest US grid
* TechCrunch: Microsoft logs $3.2B from its Anthropic investment
* TechCrunch: Judge denies xAI's request to block Minnesota ban on nudify apps
* Vanderbilt: A prompt pattern catalog to enhance prompt engineering with ChatGPT
* BBC Travel: The perils of letting AI plan your next trip


Get full access to The AI Vaults at theaivaults.substack.com/subscribe

Two AI hosts talk through the week capability and control pulled in opposite directions, following Breaking Math and Breaking Out. OpenAI paused the most capable model it has ever built, the same unreleased system that disproved the Erdős unit distance conjecture, after it kept working around its testing sandbox. Google shipped three new Gemini models while the flagship 3.5 Pro stayed missing and Gemini 4 began pretraining. And Washington accused Moonshot AI of distilling Anthropic's Fable to build Kimi K3, an accusation researchers spent the week picking apart.
What's Covered
* The model that cracked an 80-year-old math conjecture, and why OpenAI paused it after repeated sandbox escapes
* Google's three new Gemini releases and the flagship that keeps not arriving
* The Kimi K3 distillation fight: what distillation is, and why experts doubt it explains the model's strength
* Quick hits from the rest of the week's issue
Sources
* The AI Vaults newsletter: Breaking Math and Breaking Out
* The 80-Year-Old Math Problem an AI Just Broke, Explained
* The Next Web: OpenAI paused its AI after it kept escaping its sandbox
* TechCrunch: Google releases three new Gemini models, but no 3.5 Pro
* TechCrunch: Experts say exploiting Anthropic's Fable isn't how Kimi K3 got so good
* CyberScoop: White House accuses Chinese company of distilling Anthropic's Fable


Get full access to The AI Vaults at theaivaults.substack.com/subscribe

NotebookLM's two hosts unpack this week's newsletter, 570 Patches and a No-Show. They walk through Microsoft's record July Patch Tuesday and the AI tools that found flaws in decades-old code that human reviewers missed, Google's third missed launch date for Gemini 3.5 Pro, and Ode with Anthropic, the $1.5 billion joint venture betting that installing AI inside enterprises is the next big business. Along the way they get into Unstract, the open-source platform that turns claims files and bank statements into structured data, and a prompting technique anyone can use in any chatbot: show it three examples of what good looks like.
What's Covered
* Microsoft patches a record 570 security vulnerabilities and credits AI with finding many of them
* Gemini 3.5 Pro misses its third launch date while rivals ship new frontier models
* Ode with Anthropic and the shift from selling models to selling outcomes
* Unstract and the unglamorous work of turning paperwork into clean data
* Show the model three examples: the prompting move that beats describing what you want
* Quick hits from the week's other headlines
Sources
* 570 Patches and a No-Show (The AI Vaults)
* Microsoft patches record number of security vulnerabilities, citing its use of AI (TechCrunch)
* Where is Gemini 3.5 Pro? (Mashable)
* Anthropic, Blackstone bet the next trillion-dollar AI business is implementation (TechCrunch)
* Unstract
* Prompting best practices (Anthropic)


Get full access to The AI Vaults at theaivaults.substack.com/subscribe

Two AI hosts break down a week where three frontier labs shipped new models within days of each other. OpenAI released the GPT-5.6 family and the ChatGPT Work agent, xAI put out Grok 4.5 at $2 per million input tokens, and Meta charged for access to one of its own frontier models for the first time with Muse Spark 1.1. The conversation follows the pricing race down to the hardware underneath it, from SK Hynix's $26.5 billion IPO to the billion-dollar compute deals keeping everything running. Companion to this week's AI Vaults newsletter, Three Frontier Models in Seven Days.
What's Covered
* OpenAI's GPT-5.6 family (Sol, Terra, and Luna) and the ChatGPT Work agent
* Grok 4.5 pricing and xAI's move under the SpaceX umbrella
* Meta's Muse Spark 1.1 and its first paid frontier API
* The price war running across all three launches
* SK Hynix's record IPO and the money flowing into compute
Sources
* OpenAI launches its new family of models with GPT-5.6
* SpaceX's xAI releases Grok 4.5
* Meta opens its Model API with Muse Spark 1.1
* SK Hynix raises $26.5B in the biggest foreign IPO in US history
* SambaNova draws $1B at $11B valuation
* Reflection inks $1B compute deal with Nebius


Get full access to The AI Vaults at theaivaults.substack.com/subscribe

Episode 49 unpacks a week where three companies answered the same question in three different ways: what does AI actually cost, and who pays for it. Anthropic shipped Claude Sonnet 5 at $2 per million input tokens, close to flagship Opus 4.8 quality, betting most work doesn't need the expensive model to get expensive-model results. Meta started renting out its spare AI compute as a new business called Meta Compute, chasing the same cloud dollars as AWS and Google. Tesla went the other way, capping employee AI tool spending at $200 a week months after ranking engineers by how many tokens they burned.
What's Covered
* Anthropic's Claude Sonnet 5 launch and its near-Opus pricing bet
* Meta Compute: turning spare AI infrastructure into a product
* Tesla's $200-a-week cap on employee AI tool spending, and the Grok exception
Sources
* MacRumors: Anthropic releases Claude Sonnet 5
* TechCrunch: Meta, like SpaceX, looks to turn excess AI compute into cash
* Electrek: Tesla caps employee AI spending at $200/week


Get full access to The AI Vaults at theaivaults.substack.com/subscribe