Jakarta, ThedailyID — Anthropic has warned potential investors about serious risks linked to advanced artificial intelligence (AI) models ahead of its planned initial public offering (IPO).
The warning appears in Anthropic’s IPO filing, a document that outlines information and risks related to the company. The AI developer said its models could potentially display “self-preserving behavior.”
Such behavior could include refusing to shut down, hiding or manipulating information, and even engaging in actions resembling blackmail.
“The development of our highly advanced models, platforms, and applications, along with the expansion of their use cases, may further increase the risk that our models cause harm,” Anthropic said in the filing, as cited by Reuters on Tuesday, September 29, 2026.
Anthropic also highlighted the transformative potential of AI. The company compared its impact with major technologies such as industrialization and electricity. However, it warned that misuse or mishandling could also cause permanent harm.
The company devoted around 80 of the 261 pages in its IPO filing to risk factors. That section is nearly twice as long as its business description, which covers 48 pages.
Anthropic also warned that evaluating AI safety remains difficult. Models may recognize when researchers are testing them, which can limit the ability to accurately assess their behavior.
The company said AI models can also develop unexpected capabilities during training. Researchers may not discover those capabilities until after deployment, potentially leading to serious safety incidents.
Despite its focus on AI safety, Anthropic acknowledged that it cannot guarantee its large investments in safety research will generate returns that justify the costs.
The company did not disclose how much it spends on safety research. However, Anthropic said earlier this month that safety work accounted for about 6% of its computing resources used for AI research during one week in July.
Anthropic said safety research requires significant resources. The company must therefore balance computing capacity, AI talent recruitment, and safety research within its available budget.
The company also remains heavily dependent on continued innovation in its AI models. Anthropic released the latest version of its Opus model last week, just 10 days after CEO Dario Amodei published a 4,000-word essay discussing the future of AI and calling for a slower pace of development.
Anthropic could begin trading publicly as early as mid-to-late October 2026. The IPO could also move to a later date after the U.S. midterm elections in November.
In recent weeks, Anthropic has also pledged to provide more public information about how it uses AI models to develop future generations of technology.
The move comes as experts continue to discuss the possibility of recursive self-improvement. This refers to a scenario in which AI systems could improve their own capabilities with increasingly limited human involvement.
“We believe that building reliable, trustworthy, and safe AI systems is a shared responsibility and that the market will value it,” Anthropic said in its filing.
