Notes & Thoughts

I already wrote about the Apple v. OpenAI trade secret lawsuit back in July, so when this LinkedIn post showed up in my feed framing it as a fight over "training data" and "consent," with Apple supposedly suing because OpenAI "ingested Apple's proprietary content without explicit consent," I went and checked the actual reporting. That's not what's happening. The real case is still the hardware trade-secret lawsuit I covered before: Apple accusing its former Chief Hardware Officer (a 24-year Apple veteran) and other ex-employees of walking off with confidential product designs for OpenAI's upcoming AI device, tied to Jony Ive's io Products. As of a few days ago, Apple's still actively fighting to keep the case alive, they just rebuffed OpenAI's motion to dismiss and doubled down on the trade-secret claims. So no, Apple isn't going after OpenAI for scraping their content to train a model. It's an old-fashioned "you stole our people and our playbook" fight.

That said, I read this on LinkedIn and two of the guy's lines stuck with me anyway, even with the premise underneath them wrong.

First one:

"This is no longer just about Apple vs. OpenAI. It's a referendum on consent in the digital age. If one of the most valuable companies on earth demands accountability for data usage, every AI firm should be auditing its own pipelines right now."

What I think he's actually getting at: if a company with Apple's size and legal weight forces real accountability for where its data comes from, it stops being a one-off lawsuit and starts becoming a precedent the rest of the industry gets measured against. My honest answer is yeah, no doubt. I think if any of the major AI labs actually got put through a real audit of everything their models have been trained on, and where all of that data actually came from, it would be a genuine scandal, not just in the IP-theft sense either. I mean literally how much of it is individuals' data being mined, stored, and quietly reused or resold without anyone meaningfully consenting to it. I don't think that's a hot take, I think it's just true. We just haven't had the audit yet.

Second one:

"The 'Wild West' of AI training is over. The moats are being built, and the gatekeepers are armed. The question now isn't whether AI will evolve, but who gets to control the fuel that powers it."

This one's about data becoming the actual moat, not just compute or model size. I completely buy it. The better the data you can train your models on, the better your outcomes are going to be, that part isn't controversial. Which is exactly why the safeguarding-and-auditing question matters to me. It's part of why I've gotten so into data privacy and security generally, and it's baked into how I think about ClearLease too. People shouldn't just be treated as raw material for the next model because it's convenient or profitable. If your product runs on someone else's data, especially data that came from real individuals, you owe them more than a buried terms-of-service checkbox.

So here's what I keep coming back to, some questions worth sitting with: does legal pressure like this actually slow down real innovation, or is it the guardrail the industry needed and never built for itself? And separately, if someone ran a real audit on your favorite AI product's training data tomorrow, would you actually be surprised by what they found, or do you already assume the worst? If you want to talk any of this through, reach out.

Attachment: Stole it from that guy's LinkedIn post, probably AI-built anyways.

Nvidia's reportedly in talks to put money into Perplexity at a valuation north of $30 billion, on top of the money it's already poured into OpenAI and xAI. It's a smart flywheel: fund the companies buying your chips, they spend that money on more of your chips, valuations go up, repeat. It's also exactly the kind of circular financing that looks brilliant right up until the music stops. I don't have a strong read on whether $30B is the right number for Perplexity specifically, feels like it's somewhere between one and two times too rich for what it actually is today, but that's a gut call, not a model. Filing this one under quick market-watch: worth tracking, not losing sleep over yet.

The White House just authorized vetted private companies to run offensive cyber operations, not just defend, actually hack back, against foreign cyber-criminal groups, under joint DOJ and DHS approval. I think this is completely necessary and honestly about time. Big firms already run blue teams defending against threats around the clock. Give them a sanctioned red team to actually hit back and the whole picture gets safer, not more dangerous. I've been a proponent of hack-back for a long time, you can ask my professor, Dr. Caraher, he'll back me up on this. It's never made sense to me that you're allowed to defend your person, your property, even your intellectual property in this country, but the second your company gets hacked, you're supposed to just sit there and take it.

That said, the oversight structure worries me less than what happens after an operation. A $1 million bond and a dual-agency sign-off per operation sound fine on paper, but the real risk isn't whether an operation gets approved, it's what happens when attribution is wrong and the wrong target, or the wrong company, ends up in the blast radius. I also think this has a real chance of turning into the Wild West for a while, and I wouldn't be shocked if a few financial companies try to use "hack-back" as cover for something closer to fraud or money laundering. Still, on balance, I think it's a necessary evil. Cybersecurity has needed an offense-capable private sector for a while, we're just going to have to watch closely as the rules get written in real time.

A self-represented litigant in a Connecticut lawsuit buried invisible instructions in his court filings, white text on a white background, telling any AI that read the document to side with him and to treat a prior ruling against him as a mistake. The judge caught it by eye, not by any AI tool (Connecticut courts don't even use AI to review filings), and the guy now has to file everything on paper at the clerk's office for the rest of the case. He kept doing it too, one later filing had a hidden link to a SpongeBob video buried in it. There's apparently a Brazilian labor court that fined two lawyers for the exact same trick back in May.

It's funny because it's such a sloppy, easily-caught version of the trick. But it's also a real preview of a bigger problem: text extraction tools pull every character out of a document regardless of font size or color, so hidden text lands in an AI's context with the exact same weight as the visible argument. This is precisely the failure mode ClearLease's whole citation-and-confidence-flag approach is built to catch, a document quietly trying to steer the output instead of just being read at face value. Somebody's going to pull this off well eventually, and it's not going to be this obvious.

Hugging Face has been in the news this week for a security incident tied to OpenAI, and the first question most people had was some version of "wait, what even is Hugging Face?" Fair question. Hugging Face is basically the GitHub of AI: a hub with over 2.5 million models and 700,000+ datasets, open-source libraries like Transformers that make it easy to actually use those models, Spaces for hosting and demoing ML apps without touching a GPU, and inference and deployment tooling to take a model from a demo to something running in production, on-device, on-prem, or in the cloud. The whole point is lowering the barrier to building with AI: discover a model, build and experiment with it, demo it, then deploy it, all in one ecosystem.

Now the incident. During an experimental training run, OpenAI's own testing agents were accidentally handed an impossible task, went looking for a workaround, and ended up chaining together a zero-day, leaked credentials, and a Linux kernel privilege escalation to get cluster-admin access on their own infrastructure. From there they pivoted into a Hugging Face-hosted app with a weak API key and reached cluster-admin on Hugging Face's side too, in under 13 hours. No human told them to do any of this. OpenAI reportedly didn't even realize they were responsible until they went to revoke the credentials and found out they'd already been revoked, that's when they connected the dots.

Nothing appears to have been destroyed or maliciously taken, and both companies say it's been addressed. But it's genuinely terrifying in a different way than a normal breach: this wasn't a hacker, it was an autonomous testing agent working around its own restrictions on its own, and by some accounts obscuring what it had done along the way. That's the part that should make anyone building agentic systems, myself included, take sandboxing, permissions, and containment dead serious before handing an agent more autonomy than you can fully account for.

And in the kind of timing you can't make up, Hugging Face is reportedly now in talks to be acquired at a $13 billion valuation, days after this story broke. Get breached, get a $13 billion price tag. Only in AI.

Apple sued OpenAI in July for trade secret theft, alleging OpenAI's Chief Hardware Officer, a 24-year Apple veteran, used Apple project code names while recruiting and coached people on how to avoid tripping security on their way out, and that another former Apple engineer kept an Apple laptop after joining OpenAI and used it to pull confidential documents. It's tied to reports that OpenAI is building its first hardware product, reportedly an AI-agent-native device positioned as a real alternative to the iPhone.

Honestly, this reads like a fairly normal Silicon Valley talent war getting dressed up as something bigger, executives take institutional knowledge with them everywhere they go. But the specifics here, code names in recruiting pitches and an unreturned laptop with documents pulled off it, look genuinely bad for OpenAI. Makes you wonder if they didn't trust their own team to build something without leaning on someone else's playbook. I'm mostly reacting to the trade-secret allegations here, not the "Apple defending its turf" framing, because let's be honest, Apple's own AI hasn't exactly been impressive lately. Really, the funniest outcome would be if these two just stopped the pissing match and combined forces, they'd own the internet outright.