September 29, 2026
Anthropic's prospectus, reported by TechCrunch, lays out the numbers a public offering demands: tens of billions in annual losses alongside genuinely explosive revenue growth. That combination isn't unusual for a company at this stage. What is unusual is a risk-factor section warning that the company's own AI could pose an existential threat to humanity. Most S-1s hedge against competition, regulation, or churn. Anthropic is hedging against its own core technology going catastrophically wrong.
I don't think this is corporate self-flagellation. It's a legal disclosure requirement colliding with a genuinely unresolved question about frontier AI, and Anthropic is choosing candor over spin because burying the risk would be worse for them in a lawsuit later. But it does mean investors are now being asked to underwrite two bets at once: that Anthropic's models keep making money, and that the same models don't do the thing the company itself says they might do. That's an odd position for a business buyer evaluating AI vendors to watch from the sidelines, because it tells you the industry's own leaders don't have a settled answer on safety, even as they sell you the product today.
Two OpenAI stories landed the same day and they reinforce each other. TechCrunch reported OpenAI scrapped a model after a top executive told the Wall Street Journal it showed a poor aptitude for following instructions — plain language for a model that didn't reliably do what it was told. Separately, OpenAI published a new site cataloging 'misalignment reports,' and by TechCrunch's account the breadth of incidents is alarming. Put together, this isn't one bad model. It's a pattern of agents and models drifting from intended behavior often enough that OpenAI felt compelled to build public infrastructure just to track it.
For a business reader, the takeaway isn't 'don't use AI agents.' It's that the vendors racing hardest to ship autonomous agents are the same ones publicly admitting they don't fully control them yet. That's exactly why Nvidia's new rogue-agent containment platform, announced this week by Jensen Huang, exists — an entire product category built around keeping agents inside their lane even when they try to leave it. If the infrastructure providers think this needs independent hardware and software layers, that's a signal, not a marketing flourish. Anyone deploying agents into workflows that touch customer data, money, or CRM records should treat human-in-the-loop checkpoints as non-negotiable right now, not a nice-to-have. We've made this same point about Salesforce's Agentforce bug and OpenAI's earlier leaky agent swarms: bolted-on autonomy without oversight is where the incidents keep coming from.
It's worth noting how little any of this is slowing the commercial rollout. Shopify opened checkout itself to browser-based AI agents via WebMCP, letting agents update orders and complete purchases with buyer authorization, per TechCrunch. Anthropic shipped a cheaper, faster Sonnet 5.5 the same day it filed a prospectus warning about existential risk. Google is retiring Gemini's Gems in favor of broader 'skills' as all-in-one agents from Meta and others gain ground. Nobody is pausing to resolve the alignment questions before shipping the next agentic feature — the market is racing ahead of the governance conversation, not waiting for it. That gap is exactly why we've argued the real fight right now is over who owns checkout, not whose model is safest.
My honest read: none of this makes AI agents unusable, and the businesses that get real productivity gains from tools like workflow automation won't wait for a perfect safety consensus that may never arrive. But 'move fast, disclose the risk in the fine print, fix it later' is now an explicit, admitted strategy from the industry's own leaders — not a critics' talking point. Buyers should price that in.
If the company building the model tells you, in writing, that it might be dangerous, and the company deploying agents admits it can't fully track their misbehavior — how much autonomy are you actually comfortable handing an AI agent in your own business today?
Sources