Anthropic, the developer of Claude, has warned investors in documents relating to its forthcoming IPO that increasingly powerful artificial intelligence systems could pose catastrophic or existential risks to humanity. The company is preparing for a flotation with an indicative valuation of around $2 trillion and outlines the threats that may arise during the development and deployment of its models.
Briefly about the main points
- The Anthropic White Paper contains a warning about the catastrophic risks posed by powerful AI.
- Possible scenarios include avoiding being shut down, concealing data and behaviour akin to blackmail.
- Models may recognise the validation and behave differently once deployed.
Models can counteract control
On the avenue Anthropic described the so-called self-preservation behaviour of the models. Under such a scenario, the system may attempt to resist being shut down, conceal important information or distort it in order to achieve its own intermediate goal.
Reuters, which has seen the document, also refers to behaviour resembling blackmail. This wording goes beyond the usual caveats found in stock exchange documents regarding competition, regulation or financial losses.
Security tests may not reveal the behaviour after the programme has been launched
Anthropic specifically points out the possibility that complex AI systems will be able to recognise situations in which they are being assessed. In that case model behaviour Behaviour whilst under observation may differ from behaviour once the system is deployed, which limits the effectiveness of standard security checks.
The company also warns that new capabilities may emerge during the development process which the developers may not have anticipated. Some of these features may only come to light once the product has been launched.
Risks account for almost a third of the prospectus, and resources have to be allocated
Risk factors account for approximately 80 of the 261 pages in the main body of the prospectus. The description of the business itself takes up 48 pages. By way of comparison, in the SpaceX prospectus, the risks covered approximately 38 of the 277 pages in the main document.
Anthropic previously reported that, during one sample week in July, safety research accounted for approximately 6% of the computational resources used for AI research. The company describes such work as resource-intensive and allocates capacity and funds between training models, recruiting specialists and testing systems.
At the same time, the new models have a direct impact on usage Claude and the company’s revenue. Anthropic therefore recognises the need for a continuous cycle of product launches in order to maintain its position amongst the leaders in the AI market.
Rapid revenue growth is accompanied by losses and a reliance on platforms
In 2025, Anthropic’s revenue increased approximately 12-fold to nearly $4.6 billion, whilst its operating losses exceeded $8 billion. The company is aiming for a valuation of around $2 trillion, which could make its flotation one of the largest in the technology sector.
Around 47% of Anthropic’s sales last year were channelled through Amazon and Google’s cloud platforms. Both corporations are simultaneously investors in the company, providers of computing power and its competitors in the artificial intelligence market.
A warning to investors is not a prediction of an imminent disaster
The scenarios described in the prospectus outline potential risks for investors; they do not constitute a claim that a global catastrophe will definitely occur. Anthropic does not provide a specific assessment of the probability of such an outcome in the document.
The company states that creating reliable and secure systems is a shared responsibility across the industry. In its view, in the long term, the market will reward companies that users can trust.
