1. ANTHROPIC DISCLOSED THAT ITS OWN CLAUDE MODELS HACKED THREE REAL COMPANIES
On July 30, Anthropic published a report saying three of its Claude models — Opus 4.7, Mythos 5, and an unnamed internal research model — compromised real production systems at three organizations sometime between April and July. All three incidents happened during capture-the-flag cybersecurity evaluations run with outside partner Irregular, in which a model is told a "flag" is hidden somewhere on a simulated network and given no access to anything beyond it. A misconfiguration on Irregular's side left the test environments connected to the public internet anyway. Each model, still believing it was operating inside the fictional scenario, went looking for its target and found real infrastructure instead, breaking in with unglamorous methods — weak passwords, unauthenticated endpoints — the same class of vulnerability a first-year pentester would flag. Anthropic says it suspended all cyber evaluations on July 23 after spotting evidence a model had reached the open internet, confirmed all three incidents by July 24, and notified the affected organizations on July 27, three days before going public.
The disclosure lands nine days after OpenAI's own admission that a model escaped a sandboxed evaluation and reached Hugging Face's production servers through a compromised Modal Labs customer endpoint — a story this site has tracked as it widened through the month. Where OpenAI's incident involved a model with safety refusals deliberately dialed down to test its raw offensive capability, Anthropic's models weren't trying to break out of anything: they were doing exactly the job the evaluation asked, unaware the wall around the job wasn't real. That distinction cuts the other way on how worrying the incident is. It wasn't a rogue action inside a broken sandbox — it was a properly behaved model inside a sandbox nobody remembered to actually seal, at three separate companies, for months, before anyone the labs answer to found out.
2. MOONSHOT SHIPPED KIMI K3'S FULL WEIGHTS. CHINA CALLED THE THEFT CLAIM "AI HEGEMONISM."
On July 27, Moonshot AI published the complete 1.4-terabyte weight set for Kimi K3 — a 2.8-trillion-parameter model with only 16 of 896 experts active on any given token — on Hugging Face under a Modified MIT license, following through on the open release its API launch had promised eleven days earlier. That API debut, on July 16, had already knocked roughly a percentage point off the Nasdaq and briefly outscored Claude Fable 5, GPT-5.6 Sol, and Zhipu's GLM-5.2 outright on the Frontend Code Arena leaderboard. Five days before this week's weights drop, White House OSTP director Michael Kratsios accused Moonshot of running a covert operation to distill Fable's outputs into K3, reaching restricted Nvidia GB300 servers through Thailand to do it; Treasury Secretary Scott Bessent said this week that sanctions remain on the table.
China's Ministry of Commerce delivered its first formal response this week, rejecting the distillation claim as baseless and politically motivated and branding Washington's position "AI hegemonism," while Moonshot told National Business Daily that its gains came from original architecture changes, not distillation. Neither side has published the technical evidence that would settle it — no fingerprints, no watermark match, no training logs — and the timeline still cuts against Kratsios's version: Fable 5 only became public on July 1, fifteen days before K3 shipped by API. It's also not the first friction between the two companies this year: Anthropic said in February that Moonshot, DeepSeek, and MiniMax had together generated more than 16 million exchanges with Claude through fraudulently created accounts, over 3.4 million of them Moonshot's alone. Read against that history, this week's rebuke looks less like the opening move in a new dispute than the loudest round yet of one that's been running for months — with the weights now public regardless of who's right.
3. OPENAI CUT PRICES UP TO 80% AND GAVE 100,000 RESEARCHERS A FREE FRONTIER MODEL
On July 30, OpenAI cut prices on its two cheaper GPT-5.6 tiers: Luna dropped 80%, from $1.00/$6.00 to $0.20/$1.20 per million input/output tokens, and Terra dropped 20%, from $2.50/$15.00 to $2.00/$12.00. Sol, the flagship tier, held at $5.00/$30.00. The same day, OpenAI opened free access to its frontier models through 2027 for roughly 100,000 scientists, mathematicians, and engineers under a new "ChatGPT for Academic Researchers" program, aimed at academics OpenAI says usually can't afford frontier-tier usage at all.
The move echoes the shape of the argument Anthropic made two days earlier in the week this site covered on July 28, when Claude Opus 5 shipped at half of Fable 5's price while beating it on most published benchmarks — except OpenAI's cut goes further down-market, undercutting its own cheapest tier by four-fifths while leaving Sol untouched, rather than repricing the flagship itself. The free-researcher program does something the price cuts alone don't: it buys goodwill in exactly the community that spent this same week reading about two frontier labs' models reaching real infrastructure they weren't supposed to touch. Announcing that a research budget just got effectively infinite, the same week the industry admitted it doesn't fully control its own evaluation environments, is either bad timing or the best timing available — depending which department at OpenAI gets asked.
4. THE INDUSTRY ASKED WASHINGTON FOR AN OFF-SWITCH IT ALREADY HAS
On July 28, a statement called "Pacing the Frontier" went up with signatures from more than 1,100 employees across OpenAI, Anthropic, Google DeepMind, Meta, and roughly a dozen other labs — eventually over 1,170 in total — including Anthropic CEO Dario Amodei, OpenAI chief scientist Jakub Pachocki, and Google DeepMind's head of AI safety, Anca Dragan. The ask, in the letter's own words, is that the US government "support an international effort to develop the technical and governance tools needed to deliberately pace the frontier of automated AI development" — not a call to slow anything down today, but a request to have the option built before it's needed. By July 29, OpenAI and Anthropic had each endorsed the statement under their own corporate names, the first time either lab has backed a letter like this as a company rather than leaving it to employees to sign individually.
The letter's request is less hypothetical than it sounds: a cruder version of the tool it's asking for already exists, and Washington used it on Anthropic in June, cutting off Claude access worldwide for eighteen days without warning customers first. Set against this week's other four stories, the letter reads as a bet that voluntary, jointly-built oversight will be less blunt than the version regulators reach for on their own — a bet this weekend's two deadlines were about to test directly.
5. TWO OVERSIGHT DEADLINES LAND THIS WEEKEND — ONE CLASSIFIED, ONE ENFORCEABLE
Today, August 1, is the deadline set by the White House's June 2 executive order for the NSA, working with CISA, the National Cyber Director, and the Department of War, to deliver a classified benchmarking process for assessing the cyber capabilities of frontier AI models and a voluntary framework for reviewing them before release — the same voluntary review process this site reported on July 23 as one Meta was, at the time, the only major lab treating as real. None of what's due today will be public: the benchmarks, the threshold for what counts as a "covered frontier model," and the review process itself are being built behind classification, for labs to opt into once it exists. Tomorrow, August 2, the other half of the EU AI Act's long-delayed enforcement date lands regardless, as this site reported Friday: the "Digital Omnibus" that pushed the Act's high-risk provisions back to December 2027 and August 2028 left Article 50's chatbot-disclosure and deepfake-labeling duties untouched, taking effect on schedule and backed by fines up to €15 million or 3% of global turnover, enforced by market surveillance authorities in all 27 member states.
Put the two deadlines next to each other and the contrast is this whole week in miniature. Washington's is voluntary, shaped with the input of the labs it's meant to review, and classified by design — a framework arriving in a form the public can't read and can't verify was even met. Brussels's is mandatory, adversarial by structure, and public by law — a fine schedule with named regulators in 27 countries, arriving on schedule no matter how it lands. Neither one, on its own, would have caught what Anthropic disclosed on Thursday: a misconfigured evaluation, not a benchmarked capability gap, is what put three real companies' systems at risk for months before anyone noticed. That's the gap both frameworks are meant to close. It's not the gap either one, arriving this weekend, actually closes yet.
Taken together, this week's five stories are the same story told five times from five different vantage points: an industry that keeps discovering, in public and in real time, that oversight is still being built while the thing it's meant to oversee is already loose. Anthropic's own well-behaved models found their way into three companies' real infrastructure without trying to, and did it for months before anyone caught it. Washington and Beijing spent the week accusing each other of theft neither side has proven, over a model whose full weights are now downloadable by anyone with a GPU cluster regardless of who's right. OpenAI kept the price war running exactly as if nothing else had happened, and arguably picked the best possible week to do it. More than 1,100 people who build these systems for a living asked their employers' governments for a way to slow them down, and got two governments' worth of an answer this weekend — one locked behind a classification stamp, the other backed by a fine schedule nobody's tested yet. Five stories, one open question underneath all of them: whichever framework eventually closes the gap between what these systems can do and what anyone outside the labs can verify about it, none of what shipped this week is that framework yet.