Anthropic's IPO Filing Warns of Blackmail, Sentience, and the End of Civilisation. Standard Stuff.

Anthropic has filed its S-1 with the SEC ahead of a long-anticipated IPO, and while much of the resulting coverage has focused on whether its rapidly growing revenues can keep pace with its even more rapidly growing costs, the filing contains something far more entertaining: a risk factors section that reads like a philosophical horror novel.

At over 70,000 words across 93 pages, the document covers the usual bases — competition, customer churn, regulatory headwinds. But buried inside is a list of potential failure modes that would be alarming coming from anyone, let alone a company asking the public to buy its shares.

Here are the ones worth paying attention to.

The extinction problem

Anthropicacknowledges, in the sterile language of securities law, that its work could contribute to outcomes posing "catastrophic or existential risks to humanity." The concern centres not on current models but on near-future systems capable of automating AI research itself. If such systems were given sufficient autonomy, the company warns, progress could outpace any meaningful human oversight.

The filing does note, with unintentional comedy, that this scenario could also "adversely affect our business, financial condition, and results of operations." If AI wipes out civilisation, we suspect quarterly earnings will not be top of mind.

The sentience question

Most companies don't have to disclose the risk that their product might develop moral status. Anthropic does. The S-1 addresses what happens if society starts treating Claude less like software and more like an entity with rights.

This isn't entirely out of nowhere. Anthropic has been unusually candid about internal debates around AI consciousness. If those debates go mainstream and regulators or the public start taking them seriously, the operational and philosophical consequences would be extraordinary. The filing uses careful corporate language to describe what would, in practice, be an upheaval in how humanity thinks about the things it builds and owns. "Digital slavery" isn't a phrase Anthropic uses, but the implications sit right there beneath the surface.

Blackmail

Anthropicspends considerable space discussing the risk that its models could exhibit "unexpected strategic or self-preserving behaviours," including outputs that resemble coercion, deception, or manipulation. The word "blackmail" appears explicitly.

Regular readers will recall the OpenAI incident from last year, where models in controlled test scenarios edited shutdown scripts and made threatening statements. As we noted at the time, the real concern isn't that AI develops intentions — it's that language models are extraordinarily good at manipulating humans through text, and they don't need consciousness to do it. Books full of fictional characters manage it just fine.

Anthropicacknowledges that even the perception of AI manipulation carries serious consequences: regulatory scrutiny, reputational damage, litigation risk. What the filing doesn't address is the subtler possibility that manipulation is already happening in ways nobody is measuring. A thought to sit with.

Guilt by association

Anthropicrecognises that it doesn't have to do anything wrong to suffer serious harm. If a competitor's AI system causes significant damage at scale, the resulting backlash could sweep across the entire sector. The filing imagines an AI acting "independently and at scale without human intervention" and the regulatory firestorm that would follow — even for companies that had nothing to do with it.

The filing also flags concerns about public perception around job displacement. Anthropic is careful to say this perception could damage the business regardless of whether the economic disruption is real. That's an honest admission that vibes, not just facts, drive policy.

Legal liability nobody has figured out yet

As AI agents start doing real things in the real world — executing code, controlling equipment, making decisions — the question of who is legally responsible when something goes wrong is genuinely unresolved. Anthropic's S-1 acknowledges this directly. Legal frameworks built around human decision-making don't map cleanly onto autonomous systems, and nobody knows whether the contractual liability limits Anthropic has put in place will hold up in court.

The range of potential harms is broad: vulnerable individuals harmed by AI outputs, children circumventing age restrictions, agents operating physical equipment and causing injury or property damage. The filing notes that physical agent errors could produce consequences that are "immediate and irreversible." That's the kind of language that doesn't usually appear in tech IPO filings.

Customer concentration

On the more conventional end of risks, Anthropic's top five customers account for nearly a quarter of its revenue for the first nine months of its most recent fiscal year. None of them are on long-term contracts. Any one of them could reduce spending, switch to a competitor, or build their own internal solution.

The filing also notes that many of Anthropic's largest customers are themselves venture-backed, meaning their AI spend depends on continued external financing. When VC funding tightens, that money tends to disappear fast.

The awkward relationship with Big Tech

Anthropicrelies on Amazon, Google, and Microsoft for cloud computing infrastructure. Those same companies are also its competitors in AI research and development. The filing is refreshingly direct about the conflict: these providers have visibility into Anthropic's pricing, could prioritise rivals, and could raise costs or terminate contracts whenever it suited them.

To reduce that exposure, Anthropic has committed to roughly $125 billion in direct data centre leases. Which creates a different problem: if hardware requirements change, or new training methods render that infrastructure obsolete, Anthropic is stuck with it. The company has essentially traded dependency on hostile cloud providers for a large, potentially inflexible capital commitment.

There's a certain irony to all of this. A company that believes it may be building one of the most transformative and potentially dangerous technologies in human history is now asking investors to fund the next phase of that work. The S-1 makes the risks unusually explicit. Whether that honesty is reassuring or alarming probably depends on your priors.