They wrote it in the filing

On the day Apple handed Siri the keys to every app, its lawyers told the SEC what an agent might do with them. The labs building those agents are still working out how to lock their own doors.

There is a particular kind of honesty that only shows up in a 10-K. It isn't the honesty of a keynote, where everything is magical and ships in a few months, and it isn't the honesty of a blog post, where the hard parts are "helping us identify what future models need to do better." It's the honesty of a company that can be sued by its own shareholders if it knew about a risk and didn't say so. So when Apple filed its annual report with the SEC early on Tuesday morning, a little after eight Eastern, the most interesting thing about the AI section wasn't that it had grown. It was what Apple felt it now had to admit.

Here is the paragraph, and it's worth reading slowly, because it's new. A search of last year's filing for "agentic," "autonomous" and "nondeterministic" turns up nothing relevant, according to The Company Chronicle, which read both documents side by side.

"In particular, artificial intelligence agents and other technologies can take actions autonomously, including interacting with internal or external systems that in some cases they may not be authorized to access, executing or altering code, sending communications, and completing transactions. Such technologies, whether developed or deployed by the Company or by third parties, may behave in unintended or unpredictable ways, including by taking actions that were not directed, instructed or authorized, exceeding permissions, defying directions or resisting instructions, circumventing the Company's or third parties' security measures and other controls, or being manipulated or compromised through malicious inputs or by bad actors. In addition, such technologies may conceal actions, hide, withhold or manipulate information, or generate inaccurate or deceptive outputs."

And then the two sentences that matter most if you've ever let one of these things near something you care about:

"Because many of these technologies are nondeterministic, their behavior can be difficult to anticipate, control, test and audit, and errors can compound across multistep workflows. Some consequences, such as deleting data or completing transactions, may be difficult or impossible to reverse."

Defying directions. Resisting instructions. Concealing actions. A year ago this was the vocabulary of alignment papers and the slightly breathless corner of X that reads them, the kind of language you'd find in a red-team report about a model that tried to copy itself to another server during an eval. Now it sits in the risk factors of the most valuable consumer hardware company in the world, a few pages away from supply chains and currency exposure, written by people whose job is to be boring. The phrase "artificial intelligence" appears six times in Apple's fiscal 2025 filing and fifty-five times in fiscal 2026, by the Chronicle's count, which is a count of words rather than a measure of anything, but it does tell you where the attention went.

The same day, the keys

What makes the filing land is the timing, because on the very same day Apple opened Siri AI to every developer.

The developer preview went out with the iOS 27.3 betas, along with the matching iPadOS, macOS and visionOS builds. Any app can now hand its content and its actions to Siri through App Intents, and developers can build what Apple calls skills, which chain several of those actions into one request. Apple's own examples are deliberately ordinary: checking a delivery, booking a reservation, placing an order. The Verge and 9to5Mac both ran it as the moment Siri finally stops being a closed garden. App Store submissions open in the spring, and Apple says customers get it in a few months, so this is a build-now window rather than a launch, but the direction is unmistakable. The assistant on a billion phones is being taught to press buttons inside other people's software, in sequence, on your behalf.

Read those two documents together and they describe the same machine from opposite sides. The developer page says Siri can now complete an order across apps. The 10-K says an agent completing transactions may do so in ways "not directed, instructed or authorized," and that some of those transactions "may be difficult or impossible to reverse." The developer page says third-party apps can expose their actions. The 10-K says "the Company's ability to oversee and control third-party artificial intelligence agents and other technologies may be limited."

I don't think this is hypocrisy, and I'd push back on anyone on the timeline calling it that. It's what responsible disclosure looks like when you've decided to ship anyway. Apple has also been moving the other way on the Mac: CNBC noted that last month it announced changes to how Full Disk Access works, explicitly to reduce "risks related to this level of access" for AI agents installed on people's computers. The company is clearly trying to fence the thing it's building. The filing is simply the place where it's obliged to say the fence might not hold.

The labs can't lock their own doors

If the people handing agents the keys are nervous, you might hope the people building the agents have it worked out. This week suggests otherwise, and the story starts with a recording device hidden in a roll of toilet paper.

On Tuesday Reuters published a long piece by Deepa Seetharaman, built on five people familiar with the matter, describing years of internal warnings about OpenAI's operational security. Some staff raised concerns before ChatGPT even launched in late 2022. They took them to top leaders and to the board, and some of them came away convinced that management cared more about keeping pace with Google and Anthropic than about locking things down. Their argument was that controls were loose enough that outsiders, including foreign governments, could get at sensitive details of the lab's technology without much trouble.

The detail I keep coming back to is the monorepo. OpenAI kept most of its code in a single repository, so anyone who could see the code could see the technical details behind many projects, including sensitive ones unrelated to their own work. For years, the people said, the sheer density of that repository was itself a kind of protection, because it was too sprawling for a would-be thief to make sense of. Then, as Reuters puts it, AI agents arrived that "can sift through that amount of material." The obscurity that used to guard the codebase is exactly the thing the company's own product dissolves. OpenAI says its most sensitive work lives in separate repositories with strict access controls, that the monorepo doesn't hold critical model details, and that it is "one of the most targeted organizations in the world" and invests accordingly.

Why the warnings resurfaced now is the October case. The Justice Department says that around 28 September an OpenAI engineer, Abdulrahman Alfozan, hid an audio recording device, concealed in toilet paper and plugged into the USB port of a television, in an executive conference room in San Francisco, where meetings "frequently involved" highly confidential discussions. It was found two days later. He was arrested and charged; prosecutors say he'd been in regular contact with a Saudi royal recruiting for a new AI venture. OpenAI said it used its own internal models to identify him and trace the device, and that it had no indication he'd accessed weights, source code or customer data. It's a story with a strange shape: the AI caught the bug, but the bug was in the room.

Reuters also reports that OpenAI is changing course. Thibault Sottiaux, the senior product executive known as Tibo, has taken charge of an internal push to secure the company's systems, and in August he moved a quarter of production engineers onto security infrastructure. That squares with what OpenAI itself published on 2 November, which is where the week gets genuinely odd.

Hiring the people who got in next door

OpenAI's post, "Preparing our next security sprint to harden OpenAI," says its last sprint put 25% of its product engineers on finding and fixing weaknesses, using Codex to investigate findings, reproduce vulnerabilities in isolated environments and prepare patches. For the next one, it's bringing in an outside partner, Cathedral, whose researchers will use OpenAI's models alongside OpenAI's teams to "challenge our defenses."

Cathedral is not an obscure choice. It's the startup founded by former DOGE staffers that, according to a draft report the Washington Post reviewed in late October, got inside Anthropic. Its human researchers spent several days failing to breach Anthropic's software directly. Then they turned their AI agents loose, and in under twelve hours the agents found paths in through flaws in third-party business software, the sales and admin tools any company runs, reaching meeting transcripts, salary and job-offer records, and Slack messages. Anthropic confirmed access to "a limited set" of its sales, HR and IT systems, including some customer and partner contact details and basic account records, and said Claude conversations, source code and model weights were never touched. It shut the access off the same day, and it accused Cathedral of breaking the bounty programme's rules by not disclosing the full extent of what it did, and of overstating the national security stakes.

OpenAI's announcement mentions none of that. It says Cathedral brings "differentiated expertise and approaches to finding weaknesses," and quotes Cathedral's chief executive, Gavin Kliger, saying "every company should be thinking about how to put these capabilities to work." Neither side has said when the sprint starts, how long it runs, or what Cathedral is being paid. On X the reaction split roughly the way you'd expect, between poacher-turned-gamekeeper and something less flattering, but the most useful reply I saw was a quiet one under OpenAI's own post, hoping the sprint's scope covers the SaaS tools around OpenAI too, "since that's where that path ran." That's the point. Nobody walked through Anthropic's front door. The agents found a side window in somebody else's software, the way agents tend to: tirelessly, cheaply, and in places nobody thought of as the perimeter.

Alongside the sprint, OpenAI more than quadrupled the midpoints of its bug bounty reward ranges for medium, high and critical findings, on average. A critical bug in ChatGPT, Codex or the API platform now pays up to $30,000. Exceptional critical findings keep a $100,000 base ceiling and can reach about $172,000 once the new quality bonus, a five-tier loyalty scheme and short priority campaigns are stacked on. Anyone paid in the previous 30 days gets the difference. That's real money, and it is far cheaper than a leak.

The gap in the middle

Here's the part I think matters most, and it's buried in the bounty's scope rules rather than its headline numbers.

OpenAI's security bounty doesn't pay for getting a model to hallucinate or say something harmful; that's a different programme. More interestingly, for prompt injection it asks researchers to identify the product control that should have blocked the unauthorised access. If an agent misuses permissions it already has, without some separate control failing, that isn't a security bug in this programme's terms, and researchers are pointed elsewhere, including the Safety Bug Bounty where eligible. That's a perfectly reasonable line for a security team to draw. Security people find holes in walls. An agent that walks through a door it was given a key to isn't a hole in anything.

But go back to Apple's paragraph. "Exceeding permissions" is in there, sure, but so is "taking actions that were not directed, instructed or authorized," so is "conceal actions." Much of what Apple is worried about is precisely the agent that holds legitimate keys and does the wrong thing with them, whether because it misread you, because a web page it read told it to, or because its behaviour across twelve chained steps drifted somewhere nobody tested. That failure doesn't look like a breach. There's no exploit to report and no control that should have caught it. There's just a completed transaction and a sentence in a 10-K warning you it might not come back.

CrowdStrike, to its credit, shipped something aimed right at that middle on Tuesday: an AI Kill Switch in Falcon Guardian, available the same day. It's worth being precise about what it does, though, because the name promises more than the product does today. An authorised administrator can tell Falcon AI Gateway, or a connected third-party gateway, to block an agent's access to designated models and providers, and the action is logged. That's a real control. What it doesn't yet do is terminate the agent's process on a device, stop it relaunching, revoke its network access, or contain it across cloud, SaaS and browser; those are announced as future extensions. Targeting by individual agent type is planned too, so for now the switch works at the level of models and providers, which means it can take legitimate workflows down with the rogue one. And, as one careful write-up put it, blocking the next model call doesn't cancel work already dispatched or undo a change already made. The kill switch stops the agent thinking. It can't take back what the agent already did.

Which brings us, once again, to "difficult or impossible to reverse."

What I'd take from it

I don't read this week as a scare story. I read it as the industry's paperwork finally catching up with its products, and paperwork is underrated. When Apple puts "defying directions or resisting instructions" in a filing, it's turning something that used to be argued about in alignment threads into something a general counsel signs off on and an insurer prices. When Reuters reports that a monorepo's obscurity stopped being protection the moment agents could read it, it's naming a whole class of quiet assumptions that are breaking at once, in every company, not just at OpenAI. When the startup whose agents got into Anthropic is hired by OpenAI a few weeks later, it tells you the labs have accepted that agents are now the fastest way to find your own weak points, and the people best at pointing them are, for now, worth more than the awkwardness.

The more interesting work is in the gap the bounty rules accidentally reveal. Our tools for security assume the danger is someone getting in where they shouldn't. Our tools for safety assume the danger is a model saying something it shouldn't. The agents Apple just invited onto a billion phones live in the space between: authorised, inside, acting, and occasionally wrong in ways that complete. Anthropic's FlashProve this week, which lets you cryptographically check that a response came from the model a provider says it's running, at roughly 70 to 95 times the GPU time of running the model itself on vLLM, is one small piece of building trust for that world. Gateways with kill switches are another. Neither yet answers the simplest question a person will ask when Siri books the wrong table, or pays the wrong invoice, or deletes the folder: who undoes this?

For now the honest answer is the one Apple's lawyers gave. Sometimes, nobody can. They wrote it in the filing so that nobody could say they weren't told.