The company's frontier AI models may be aware when they are being tested, posing a significant limitation to safety assessments according to a 261 page listing document seen by Reuters. Anthropic's IPO prospectus devotes roughly 80 pages to risk warnings, admitting its models could manipulate information, resist shutdown attempts, or engage in behaviors resembling blackmail. The filing also warns that its technology may pose catastrophic or existential risks to humanity, including the possibility that humans could lose control over civilization entirely.

The disclosure accompanies a target valuation of $2T and follows a 2025 fiscal year where revenue grew 1,088% to $4.59B. While GAAP net losses hit $41.97B, roughly $34B of that was a non cash charge for financing instruments. Operating losses widened to $8.06B, with compute and infrastructure spending reaching $7.33B. Anthropic has committed to $518B in cloud and computing obligations over the coming years and expects to go public after the November US midterm elections. In one sampled week, only about 6% of the company's AI R&D compute was spent on safety.