"LOL, I found out I can access the [network storage], so funny." Apple's complaint against OpenAI names the engineer, the authentication bug, and the 400 hires, a map of what is worth stealing once the model is a commodity
The Apple v OpenAI complaint gets its first close reading, and the documents name the mechanics: an ex-engineer's "LOL, I found out I can access the [network storage], so funny" message, a previously unknown authentication bug, Apple parts brought to OpenAI interviews for "show and tell," a circulated guide to dodging the security walkout, 400+ ex-Apple employees, and io's alleged metal-finishing know-how. What's worth stealing when the model is a commodity. Beside it, the price sheet inverts: Databricks' cost-per-completed-task benchmark puts Sonnet 5 above Opus 4.8 despite 1.7x cheaper tokens, and Bun's $165,000, 11-day Claude rewrite of 500,000 lines draws "unreviewed slop" from Zig's creator. Plus Clawk's disposable agent VMs and Cloudflare Precursor.
L-O-L, I found out I can access the network storage. So funny. It's Tuesday, July fourteenth, and that text message is now evidence in Apple versus OpenAI.
The complaint reads like a spy novel. An engineer, eight years at Apple, allegedly used a previously unknown authentication bug to get back into Apple's internal network after leaving for OpenAI, and messaged a former colleague that it was, quote, so funny. Hours after his exit: I still have another computer.
Apple says the retail theft sat on top of a wholesale system. Candidates coached to bring actual hardware parts to interviews. A circulated guide to dodging the security walkout. Four hundred former Apple employees now at OpenAI.
Read the exhibit list as a price signal. Nobody stole model weights. What allegedly moved was metal finishing, supplier lists, and design process. The things that stay scarce when the model is a commodity.
And the price sheet lied again. Databricks priced models per completed task. Sonnet 5, with tokens one point seven times cheaper, costs more per finished job than Opus, because it finishes fewer jobs. And Bun's five-hundred-thousand-line rewrite by Claude agents ran eleven days and about a hundred and sixty-five thousand dollars. Code Zig's creator calls unreviewed slop.
Cheap tokens, dear tasks. The bill arrives per finished job now.
To the tape. Micron long, held. Nvidia watch, held. And we hold the Apple watch, because a company litigates this hard over metal finishing only when it believes the device layer is the moat.
The tape is the desk's scorecard, not advice.
Our call: within six months, a major provider ships a production tier priced on completed work, not tokens.
What proves us wrong is January with every price sheet still quoting tokens alone. The unit of account is moving. Watch for the first vendor confident enough to sell a finished job.
!["LOL, I found out I can access the [network storage], so funny." Apple's complaint against OpenAI names the engineer, the authentication bug, and the 400 hires — a map of what is worth stealing once the model is a commodity](https://qwrydntnyuqfhmmwdtwc.supabase.co/storage/v1/object/public/briefing-images/briefing-hero-2026-07-14.jpg)
"LOL, I found out I can access the [network storage], so funny." Apple's complaint against OpenAI names the engineer, the authentication bug, and the 400 hires — a map of what is worth stealing once the model is a commodity
"LOL, I found out I can access the [network storage], so funny." Court documents do what press releases never do: they name the mechanics. In the complaint Apple filed against OpenAI, now getting its first close public reading, the mechanics run like this. Chang Liu, eight years a senior systems electrical engineer at Apple, leaves for OpenAI in January and, per Apple's telling, exploits a rare, previously unknown authentication bug to reach the company's internal storage after his employment ends. He sends a former colleague the message above. Hours after his exit he adds, "I still have another computer." Apple says dozens of confidential hardware files about unannounced products left with him.
The complaint's larger claim is that the retail theft sat on top of a wholesale system. Apple alleges OpenAI's hardware chief Tang Yew Tan, himself Apple's former iPhone design lead, ran recruiting as acquisition: candidates encouraged to bring "actual parts" to interviews for "show and tell," asked for CAD files and prototypes, one candidate reportedly surprised to learn Apple parts could leave the building at all. Departing employees, Apple says, passed around the company's own internal "Need to know" security document repurposed as a manual for dodging the "dreaded walkout," the escorted exit that ends access the moment you resign. More than four hundred former Apple employees now work at OpenAI, and io, the hardware unit OpenAI bought for $6.5 billion, stands accused of arriving with Apple's metal-finishing know-how in its luggage.
Read the exhibit list as a price signal. Nothing in it is model weights, evals, or training recipes. What OpenAI allegedly took — and what Apple is burning its most aggressive litigation in years to protect — is metal finishing, supplier relationships, industrial-design process, and the people who carry them. On Saturday we wrote that this suit marked the fight moving off the model. The documents go further and name what the fight is over. When frontier capability is a dropdown and a price sheet, the secrets worth stealing are the ones about making a physical object beautiful, a hundred million times, at a margin.
Now the fair reading for the defense, because complaints are advocacy, not findings. Apple's version is built to read like a spy novel, and some of it may not survive discovery. Hiring a rival's engineers is legal, California will not enforce a non-compete, and every company in this industry recruits from the best team in the category it is entering. The line the case will actually test is narrower and more interesting: whether a pattern — coached exits, parts at interviews, code names in recruiting pitches — adds up to misappropriation even where each individual hire is lawful. That question, not the LOL message, is what every AI lab's counsel is reading this morning.
One more reading, and it is this week's. Strip the drama and the initial breach is an identity failure: a departed employee, an unrevoked path in, an authentication bug, internal storage reachable from outside. That is GitLost with a human in the loop — access outliving intent. The companies studying this case for hiring guidance should also file it as a security post-mortem, because the most sophisticated adversary in Apple's telling needed no model at all, just a credential nobody killed. Apple invented the walkout precisely because access does not expire on goodwill. The industry deploying autonomous agents by the thousand might sit with that for a minute.
Databricks priced models per completed task, and the cheap-token story inverted: Sonnet 5 costs more per finished job than Opus 4.8
Databricks ran real tasks from its own engineers through the leading coding models, with the evaluation overseen by CTO Matei Zaharia, and measured the unit that matters: dollars per completed task. Opus 4.8 came in at $1.94. Sonnet 5, priced 1.7 times cheaper per token, cost $2.09 — it finished 81 percent of tasks against Opus's 87, and the retries ate the discount. GLM 5.2 tied Opus on quality at $1.28. A March study found the same inversion in roughly a third of model pairings; Gemini 3 Flash lists 80 percent below GPT-5.4 and costs 38 percent more in practice. The harness compounds it: Armin Ronacher's minimal Pi hit equivalent success at half the cost of vendor harnesses, 237,000 tokens per task against Claude Code's 742,000. The price sheet quotes tokens. The bill arrives per finished job, and the two have quietly stopped correlating.
Bun rewrote 500,000 lines from Zig to Rust with Claude agents in 11 days, for about $165,000 — and Zig's creator calls the result "unreviewed slop"
The clearest receipt yet for agent labor at scale: Jarred Sumner says Bun's conversion out of Zig ran eleven days and roughly $165,000 in API costs, half a million lines rewritten by Claude agents. Andrew Kelley, who created Zig, answered that the language was never the problem — Bun was "How Not To Write Zig Code" long before any model touched it — and that shipping what he calls a million lines of unreviewed slop is the actual sin: a test suite that missed the old bugs will not catch the new ones. Zig, for its part, rejects AI-generated contributions upstream by policy. Mitchell Hashimoto praised the speed, and the speed is real. So is the question under the invoice: $165,000 bought code no human has read, and the completed-task arithmetic one story up has no column yet for what review costs.
Clawk gives coding agents a disposable Linux VM instead of your laptop
A Show HN with the week's most defensible pitch: run the agent, but not on the machine that holds your keys. Clawk hands each coding agent a burnable Linux VM — real compute, none of your laptop's credentials, browser sessions, or dotfiles — so the blast radius of a hijacked or merely overeager agent is a machine you can delete. A week after GitLost showed an agent's context window doubling as its attack surface, and the same morning Apple's complaint shows what one unrevoked credential costs, the containment tooling is arriving as weekend projects. Least privilege for machines that act is turning into plumbing you can install rather than a policy document you can ignore.
Cloudflare ships Precursor: session-long behavioral analysis to tell agents from humans
Cloudflare introduced Precursor, a client-side verification layer that watches behavioral signals across an entire session rather than at a checkpoint, on the logic that automation can imitate a person for a moment but rarely for a full visit. It slots into the Bot Management tier, evaluates aggregate patterns rather than storing individual histories, and names its target plainly: agentic traffic. The web is re-arming for a population of visitors that passes every one-time test. For anyone shipping agents that browse, the defenses are moving from "prove you are human once" to "behave like one continuously," which is far harder to fake and far easier to trip by accident. Expect your own automations to start hitting it, and expect "verified agent" lanes to become a product soon after.
Two ledgers landed today. Apple's complaint prices what stays scarce when models commoditize: process, people, and the credential nobody revoked — the theft it alleges needed no AI at all, just access that outlived employment. Databricks priced the other side: the cheap model that finishes 81 percent of jobs costs more than the dear one that finishes 87, and a $165,000 agent rewrite is a bargain right up until someone has to review half a million lines nobody read. The token price is now the least informative number in an AI budget. Count completed tasks, count review hours, count who still holds a key. The industry spent a month celebrating the first column; today's news happened almost entirely in the other three.
Within six months, a major model provider or coding-agent vendor ships a generally available production tier priced on completed work rather than tokens — per finished task, or with fees contingent on a defined completion standard. The unit of account moves from the token to the job.
Databricks just published the inversion in public: cheaper tokens, dearer tasks, in a third of pairings. Buyers are learning to benchmark per completed job, harness overhead runs to hundreds of thousands of tokens per task, and a single agent rewrite can invoice $165,000 with no completion guarantee attached. The vendors already measure task success internally — they train on it — so the first one to sell certainty, a job finished or a fee reduced, converts the market's new math into the premium tier. Pricing follows measurement, and the measurement just changed underneath the price sheets.
If, by January 14, 2027, no major model provider or agent vendor offers a generally available production tier priced per completed task or with completion-contingent fees, and production price sheets still quote tokens alone, the call is wrong.
We hold the Micron long, unchanged. Nothing on today's wire touches memory supply or pricing; the thesis — the buildout's scarce, repricing input is the memory beside the accelerator — and its named reversal risk both stand exactly where the weekend left them.
HBM and leading-edge DRAM stay tight into 2027 on AI capacity demand, with pricing power at the seller. The standing offset: memory over-corrects on a three-year fab lag, and the boom-bust case is public and dated.
We hold the Nvidia watch, no change to the question: how much of near-term demand is genuine end-use versus vendor-financed neoclouds buying with money Nvidia helped provide. Today's completed-task pricing story cuts both ways for the demand picture — retries burn more tokens, which is more compute, but buyers optimizing cost-per-job will also route harder toward whatever silicon serves it cheapest.
Real agentic demand keeps rising; a meaningful slice of near-term orders still runs through leveraged neoclouds Nvidia helps fund, which flatters the signal until the financing is tested.
We hold the Apple watch opened Saturday, and the documents sharpen it. A company litigates this aggressively over metal finishing and supplier lists only when it believes the device layer is the moat — and Apple's quiet product news this week, an on-device speech API outscoring the cloud default in a public benchmark, states the same thesis in the positive. No frontier model, full ownership of the two layers that do not commoditize: the device and the distribution. The suit is the defensive half; iOS 27's on-device stack is the offensive half.
If the AI fight settles onto hardware and distribution, Apple holds both and is now visibly defending them in court and shipping against them in the OS. The offset: it still rents its frontier intelligence, and a shipped OpenAI device that lands would test the moat directly.