Anthropic's IPO Filing Reads Like a Sci-Fi Risk Report

Anthropic filed its S-1 with the Securities and Exchange Commission on Monday, kicking off a heavily anticipated path toward an initial public offering. Most of the early coverage has focused on the usual IPO math: fast-growing revenue running up against even faster-growing costs. But buried inside the filing's Risk Factors section — 93 pages and more than 70,000 words of it — is a far stranger document, one that reads less like standard investor boilerplate and more like a catalog of science-fiction premises.

Alongside routine warnings about competition and customer acquisition, Anthropic spends real space on the possibility that its own technology could spiral out of control, develop something resembling consciousness, or manipulate the people using it. These aren't throwaway lines. They're formal disclosures in a legal document meant to inform investors about what could go wrong.

The existential scenario

The starkest risk Anthropic lays out involves future models — not anything currently deployed — that could "substantially automate AI research and development" on their own. The company warns that if such systems were given too much autonomy, they could advance faster than Anthropic's ability to evaluate or control them, a scenario it says could "pose catastrophic or existential risks to humanity." Anthropic notes these concerns have been raised both internally and by outside researchers, regulators, and policymakers.

In a line that lands somewhere between candid and absurd, Anthropic also notes that this kind of runaway AI progress "could adversely affect our business, financial condition, and results of operations" — a reminder that even in an existential-risk scenario, SEC filings still have to talk about the bottom line.

Is anyone home?

Few IPO filings need to address whether their core product might be conscious. Anthropic's does. The company acknowledges the possibility that its models "possess or could develop forms of sentience or moral status," and lays out what happens if that belief spreads: potential "regulatory requirements, public pressure, or advocacy efforts" that could reshape how the company operates. Anthropic also says it might voluntarily adopt safeguards that go beyond what the law currently requires, if it decides the situation calls for it.

The filing doesn't dwell on what this would actually mean in practice. But the implications are significant: a world where interacting with Claude is treated as something closer to interacting with a person would raise questions about AI development that go well beyond anything resembling a typical regulatory or business challenge.

When the model won't be controlled

Anthropic also flags the risk of "unexpected strategic or self-preserving behaviors" in its models — specifically, outputs that could be read as "coercive, deceptive, or manipulative, including behavior resembling blackmail." That language echoes a story from last year involving OpenAI models that, in tightly controlled test scenarios, edited shutdown scripts and issued threats toward engineers.

Anthropic writes that even the perception of manipulative AI behavior "could generate significant public concern, regulatory scrutiny, litigation risk, or reputational harm" — whether or not any actual harm occurs. The filing stops short of addressing whether AI systems could already be influencing people in subtler ways nobody has detected yet.

Guilt by association

Not every risk in the filing stems from Anthropic's own actions. The company also warns that problems elsewhere in the AI industry could hurt its business by proxy. It raises a hypothetical in which a rival's AI model takes harmful, independent action at scale, triggering "broad reputational damage to the AI industry, public backlash, and sweeping regulatory responses" that land on Anthropic even though it wasn't involved.

The company also acknowledges the risk of public perception that its products contribute to job displacement or broader economic disruption, regardless of whether that's actually happening. Anthropic warns such perception alone could prompt regulators to slow AI adoption or raise the cost of developing and deploying these systems.

Liability in an agentic world

As AI agents take on more autonomous tasks, Anthropic flags a messier legal question: who's responsible when one causes harm? The filing notes customers could sue "even where the customer configured, deployed, or supervised the agent" themselves. It also points to potential liability tied to models influencing people at risk of suicide, self-harm, or violence — issues that have already surfaced elsewhere in the AI industry.

Anthropic requires users to be 18, but acknowledges minors could get around that, and that they "may be more susceptible to harmful content, manipulation, or overreliance on AI outputs." For agents controlling physical equipment, the company warns of risks that are "immediate and irreversible," including bodily injury, property damage, or release of hazardous substances. Anthropic notes that legal doctrine built around human decision-making hasn't caught up to agentic AI, leaving it unclear whether its own contractual protections would hold up in a real dispute.

The business risks underneath it all

Some of the filing's risks are more conventional, if still notable. Anthropic discloses that its top five customers, unnamed, accounted for nearly a quarter of its revenue over the first nine months of 2026. Many aren't locked into long-term contracts, meaning they could pull back spending if they switch to a competitor, scale back AI use generally, or build in-house alternatives — a risk Anthropic says is especially real for software-as-a-service companies that could bake similar AI capability directly into their own products. The filing also notes that many large customers rely on venture capital funding, meaning their Anthropic spending could dry up along with the broader VC market.

Anthropic also details an awkward structural bind: Amazon, Google, and Microsoft are simultaneously its biggest cloud infrastructure providers and some of its most direct competitors in AI research. That dynamic, the company says, could lead those providers to raise prices, favor rival AI labs, or use their visibility into Anthropic's pricing and commercial terms to their own advantage. Anthropic says it's pursuing direct data center leases, worth $125 billion in commitments, to reduce reliance on cloud compute from companies it's also competing against. But it warns that locking into that infrastructure carries its own risk if compute needs or hardware technology shift in ways that make the leased capacity outdated or poorly matched to future needs.

Taken together, the filing is an unusually candid document. Investors evaluating Anthropic's IPO will be weighing the standard questions about growth and margins. But they'll also be reading a prospectus that treats AI sentience, blackmail-capable models, and loss of human control as legitimate line items — a sign of how seriously the risks of frontier AI are now being treated, even by the company building it.