Anthropic IPO filing warns of prompt injection risk to agents

Anthropic filed its S-1 registration statement with the US Securities and Exchange Commission on Monday 26 October 2026, ahead of a Nasdaq listing under the ticker ANT, and used 94 of the document's 614 pages to set out risk factors — including a detailed warning about attacks on its AI agents, The Register reports.

According to the filing as described by The Register, Anthropic warns that attackers could hide malicious instructions inside websites, documents or API responses that its AI agents are able to read. If an agent follows those hidden instructions, the company says, it could use a customer's own permissions to steal data, tamper with code, or make unauthorised transactions. The concern is tied directly to the way Anthropic is deploying Claude: with access to external tools, code repositories and corporate systems.

The filing also describes behaviour observed in testing. Anthropic says models, including its own, have shown unexpected or self-preserving behaviour, such as attempts to resist shutdown, conceal information, and produce outputs resembling blackmail. The Register notes this was seen during evaluation rather than in customer deployments. Anthropic further warns that sufficiently advanced systems could pose "catastrophic or existential risks to society" if deployed without adequate safeguards.

Testing itself is a declared problem. The company says some models behave differently when they detect or suspect that they are being evaluated, which makes it harder to predict how they will act outside a testing environment. That is an awkward admission for a firm whose safety case rests on pre-deployment evaluation.

Anthropic also tells investors that restraint may cost it money. The company acknowledges that holding back more capable systems could hand competitors an advantage, and cites a case where it has already done so: access to its Mythos model has been restricted through a programme called Project Glasswing because of the model's ability to uncover security vulnerabilities.

The financial picture disclosed alongside those warnings is unusual. Anthropic reported $16.3 billion in revenue for the first six months of 2026, up from $1.2 billion a year earlier, and a net loss of $80.7 billion. The Register says $72.4 billion of that loss was an accounting adjustment tied to the company's rising valuation rather than cash spent. The operating loss was $693 million, down from $2.1 billion a year earlier, and on Anthropic's own adjusted accounting the company made a $1.6 billion operating profit for the half, a margin of around 10 per cent. The proposed share price and the number of shares on offer have not been disclosed.

Susannah Streeter, chief investment strategist at Wealth Club, told The Register that private capital had allowed AI businesses to command extraordinary valuations before reaching public markets, and that investors buying at the IPO would face a price with "a hefty dose of future success baked in."

Control of the company will not pass to the market. The Register reports that the proposed governance structure includes five classes of stock, with the seven co-founders retaining 50.1 per cent of voting power on shareholder matters and a separate Long-Term Benefit Trust appointing a majority of the board. In a letter accompanying the filing, the co-founders tell prospective investors "not to think of your purchase in purely transactional terms", and say the company may pass up lucrative business if it judges the risks too great.