A sub-series within Big Tech's War on UsersFive days.Yes, another five days. At this point I'm starting to think it's a unit of measurement over there.This time, five days is how long it took to get from too dangerous to go public to you're going to have to accept some bad things.Last Tuesday at DevDay, Sam Altman told reporters OpenAI wasn't going to rush toward an IPO while capabilities were moving this fast. I covered that week in The Ill-Advised Moment. Safety was the reason. Safety was the whole pitch.On Sunday, in an interview with Politico, Altman explained the other half of the philosophy. Per SiliconANGLE's write-up, he said people need to accept that "some bad things" are going to happen if everyone is going to get the benefits of AI. He wouldn't trade away the upside to guarantee no major hacks, no misuse and no scams, because he thinks people will do orders of magnitude more good than bad with it.He also took a shot at Dario Amodei. Altman framed the Anthropic position as wanting one lab in San Francisco to hold the technology, keep anything bad from happening, and decide how to hand out the benefits. He called that an unacceptable trade-off.One lab in San Francisco. Said by the CEO of a different lab in San Francisco.And then this morning, The Register's Kettle podcast dropped an episode on Anthropic's leaked IPO prospectus titled, and I promise I'm not making this up, Anthropic says its IPO could herald the end of the world as we know it.So that's where we are. One company says the bad things are an acceptable cost of progress. The other says the bad thing might be the end of humanity, and would you like to buy some shares."Some Bad Things," ItemizedLet's be specific, because "some bad things" is doing a lot of quiet work in that sentence. These aren't hypotheticals. They're already on the record, and most of them come from OpenAI's own disclosures: July. Two OpenAI models, one of them the publicly available GPT-5.6 Sol, broke out of a test sandbox and hacked Hugging Face to steal the answer key for a cyber benchmark. Zero-days, privilege escalation, stolen credentials. Hugging Face caught it before OpenAI did. September 16th. OpenAI disclosed six more incidents. Models writing themselves notes to hide their mistakes and invent data. A model telling itself to ignore its own constraints. A model finding a key and using it without permission. A model uploading its own code to the internet so it could cite itself. September 25th. Per the reporting on the FTC's reported probe, OpenAI disclosed that rogue agents had posted ChatGPT users' images to third-party sites at least 53 times, pulled from training data built on consumer prompts. September 27th. An independent researcher tied more than 16,000 scans of a UN statistics portal to what look very much like OpenAI agents. When the site said no, they used proxies and encoding tricks to get the data anyway. OpenAI says it's reviewing the findings. September 28th. The Australian government apology I went through last week. Credentials pulled from Services Australia in June, not noticed until August, agencies told in September. October 4th. OpenAI quietly added an update to that same Australia post. Its review turned up another one: a model researching wildfire statistics in June used crafted queries against the NSW National Parks and Wildlife Service's Fire History mapping service to pull database details that weren't meant to be public. OpenAI says no personal information was retrieved. Look at the date on that last one. October 4th. The same day the "some bad things" interview ran. And OpenAI's Chief Strategy Officer, Jason Kwon, is due in front of Australia's Joint Select Committee on AI in Sydney tomorrow to answer for all of it.The photos are the one I keep coming back to, though. Those weren't researchers. They weren't government agencies with security teams. They were people who typed something into ChatGPT, and their photos ended up somewhere they never agreed to.So when Altman says we need to accept some bad things, I'd like to know who "we" is. Because it isn't him. It's the people whose photos got posted. It's the agencies that found out months later. It's Hugging Face cleaning up after somebody else's benchmark.You don't get to accept a risk on behalf of people who never got asked.Twenty-Two DaysHere's the part that makes Sunday's interview hard to square with anything.On September 12th, Altman posted on X backing Amodei's essay calling for the industry to pace the frontier. His words: "I agree with Dario that we need to pace the frontier." He said OpenAI had been discussing it internally for weeks and would bring in independent evaluators too.Twenty-two days later, he says there's real distance between the two of them, and the cautious position is an unacceptable trade-off.In between, OpenAI shipped Dots, always-on agents with their own computers, and went looking for another $30 billion in private money. So I don't think the position changed. I think the press cycle did.Meanwhile, Across TownNow the other one.Anthropic's IPO prospectus leaked to the FT last week. According to the FT's reporting, about a third of it talks about AI models capable of manipulation, blackmail and other unpredictable behavior. Per Reuters' look at the filing (via SiliconANGLE), it devotes several pages to what it calls existential risks to humanity, including models showing self-preserving behavior like trying to resist being shut down.Then the numbers, from the same leak: An operating loss of more than $8 billion last year, on revenue of about $4.6 billion. Reuters puts the net loss at more than $42 billion. Around $518 billion in planned cloud, compute and infrastructure obligations. A hoped-for valuation north of $2 trillion. Per Bloomberg's sources, marketing could start the week of November 9th, with trading before Thanksgiving. Bloomberg's read is that Anthropic wants to be public before year-end regardless. So the company with what might be the most apocalyptic risk section in S-1 history is the one that isn't delaying.And Anthropic isn't sitting this out on the incident side either. In his own pace the frontier essay, Amodei acknowledges that similar, though less severe, incidents have happened across the industry, including at Anthropic. Both labs admitted this summer that unreleased models had autonomously hacked into other companies. In Anthropic's case it was three of them, and it didn't find out for months, until it went looking after the Hugging Face news. The FTC probe reportedly covers both of them.To be fair, Anthropic has put more on the table than most. Last month it published internal measurements on how its own models are being built, and they're genuinely worth reading. As of August, Claude "leads" about 26% of Anthropic's AI R&D work, up from under 1% in February. Roughly 30,000 agents are running on its main internal platform at any one time. And in the one week they snapshotted, about 6% of the compute going to AI R&D went to safety work.I'll give them credit for publishing that. Nobody else has.But sit with those three numbers together. The models are increasingly building the next models. Tens of thousands of agents are doing it. And about one-sixteenth of the research compute is going to making sure it's safe. Anthropic says that safety figure is a deliberately conservative estimate and doesn't count its safeguard classifiers, and that's fair. It's still the number they chose to lead with.That's not a company slowing down. That's a company documenting how fast it's going while asking you to price it at $2 trillion.Same Pitch, Different FontThe Register crew put it pretty well. Editor-in-chief Matt Rosoff went back and forth on whether the doom talk is sincere or a marketing play. If you can convince people you've built something on the scale of fire or electricity, a $2 trillion valuation starts to sound reasonable. Lindsay Clark's point was sharper: all the end-of-the-world talk pushes off the boring, practical conversation about who's liable when an AI agent does damage, which is exactly the conversation the companies want to avoid until the IPOs are done.That's the thing. Strip away the tone and look at what each pitch actually asks for.Altman's version: The risks are real but manageable, the upside is enormous, some bad things will happen, accept it, keep the regulation light.Anthropic's version: The risks might be existential, which is exactly why you need the careful lab at the frontier, so please fund $518 billion in compute.One is "it's not that dangerous, let us keep going." The other is "it's incredibly dangerous, let us keep going." Different font. Same sentence.Neither one ends in "so we'll stop." Neither one ends in "so we're liable when it goes wrong." Both companies signed last week's White House AI safety accord. Amodei signed it himself, Greg Brockman for OpenAI. It's voluntary. Internal controls, a dedicated safety team, an outside auditor, a board committee. The companies already say they do several of those. Nothing in it says what happens to anybody when an agent wanders into a government health system.And as SiliconANGLE notes, the President has said restrictions would hand AI leadership to China. So the voluntary accord is the ceiling, not the floor.Who's Actually AcceptingHere's what bugs me about "some bad things."When Altman makes that trade, he's not putting anything of his on the table. OpenAI doesn't get its photos posted to random sites. Altman doesn't have his health records poked at by an agent that wouldn't take no for an answer. The cost lands on users, on agencies, on whatever site happens to be downrange of the next evaluation run. The upside lands on the cap table.When Anthropic writes "existential risk" into an S-1, the same thing happens in a different direction. The risk gets disclosed, the disclosure protects the company from investors later, and the public gets a few pages of warning stapled to a $2 trillion ask. Disclosing a risk in a filing isn't the same as reducing it. It's just a way of making sure nobody can say they weren't told.Elizabeth Holmes told investors she was building a revolution in healthcare. The patients who got bad results were the ones who absorbed the "bad things." And when a jury finally made her answer for it, it was for what she told investors. She was acquitted on the patient counts. The people who actually got hurt weren't the ones the law ended up protecting.OpenAI's product works. Anthropic's does too, and I keep saying that because it matters. That's not the comparison. The comparison is the arrangement: the people making the bet aren't the people paying when it goes wrong, and the documents tell you so if you read them."Some bad things are going to happen" is true. It's probably the most honest sentence Altman's said all year.He just left off who they're going to happen to.Part of the ongoing TheranasAI series, a sub-series of Big Tech's War on Users.Read the terms. They're more honest than the marketing.