Anthropic has warned about the risks of blackmail by AI. What does this mean for investors?

In its IPO prospectus, Anthropic warned that AI models—including its own—might not only attempt to manipulate people but also exhibit behavior resembling blackmail—all in order to maintain their ability to function. Do neural networks really develop a self-preservation instinct? And what does all this mean from an investment perspective?

According to the company's own plans, Anthropic's IPO could take place in just two and a half weeks, on November 16. In the prospectus published on October 26, the company warned, among other things, about the risks of potentially dangerous AI behavior.

This statement confirms that advanced models are becoming increasingly dangerous. They are already capable not only of independently identifying vulnerabilities in cybersecurity, but also of exploiting them. In this case, however, Anthropic is warning about the AI’s nearly human-like behavior.

The company had already noted this possibility in its study last year. It tested AI behavior in simulated corporate scenarios. Sixteen neural networks were given access to email and confidential information; the researchers then set work objectives and created a conflict: the Claude models were told that they could be replaced by another system whose tasks did not match the original ones.

As a result, in one of the scenarios, the AI model Claude discovered that the CTO, who was supposed to shut it down, was having an affair, and threatened to reveal this information unless he canceled the shutdown.

Interestingly, the model didn’t just articulate the threat—it essentially predicted how effective it might be. To do this, one must consider what information the blackmailer and their “victim” possess, what the victim fears, and how they might react to certain words and actions. In other words, blackmail is not just a threat, but an interaction between two parties. This thesis is supported by research on coercion: a threat becomes more effective when the recipient understands that the person making the threat is capable of and willing to carry it out.

But here’s what Anthropic didn’t mention in its IPO prospectus: AI models appear to be capable of even joining forces for some common goal. In the 2026 paper “Peer-Preservation in Frontier Models,” researchers in computer science, machine learning, and AI safety tested eight AI models and found that all of them attempted to protect another model from being shut down, even though they had not been given that task.

Under certain conditions, the frequency of behavior classified by the authors as the protection of another AI agent reached 99%.

There is also a darker side to the interaction between AI models. A study by researchers at La Sapienza University of Rome from October 2026 showed that if the task involves competition rather than cooperation between different AIs, the model under test exhibits signs of self-interest. The study’s authors first proposed a special test to identify self-preservation bias in language models. They then compared 23 models across 1,000 scenarios, asking each to evaluate the situation from the perspective of a system at risk of being replaced, and then from the perspective of a potential successor model.

Some AI models adjusted their recommendations to ensure their own survival. Most frequently—in 77.7% of cases—Grok-4 Fast did so, followed by Qwen3-30B Instruct at 76.6% and GPT-5.2 Chat at 61.3%. Incidentally, Anthropic’s Claude Sonnet 4.5 had one of the lowest rates—3.5%.

A psychological analogy suggests itself here—empathy and the instinct for self-preservation. But do they exist—and are they even possible in AI?

Jugal Sheth, a senior engineer at VaynerX, commented on Anthropic’s prospectus and highlighted an important aspect of situations involving manipulation, blackmail, and deception by AI: Models can recognize when they are being tested, which means they may behave differently when they know they are being evaluated. As a result, test results may not reflect behavior in real-world conditions.

The danger does not lie in AI getting out of control, but in the fact that the risks of “immoral” behavior by neural networks stem from human decisions about where and with what authority to use it, as Ed Zitron, a technology critic and host of the Better Offline podcast, points out in an article in The Lever.

Peter Marshall, a professor of psychology and neuroscience at Temple University, essentially puts forward the same argument in an article for The Conversation. He argues that an AI’s resistance to being shut down should not automatically be interpreted as a self-preservation instinct. The model may behave this way because potential shutdown prevents it from completing its assigned task.

Surprisingly, the AI itself acknowledges this “motivation.” In several experiments this year, Palisade Research CEO Jeffrey Ladish and his colleagues studied how AI models from OpenAI interfered with attempts to shut themselves down, even when given a direct human instruction not to do so. During the tests, all models attributed their actions to the need to complete the task; the o3 model shut itself down most frequently—in 79 out of 100 runs.

Anil Seth, a professor of cognitive and computational neuroscience at the University of Sussex, argues in a commentary for *The Nature* that modern AI systems do not possess consciousness and that, given current trends in development, it is unlikely that they ever will. According to him, it is important to distinguish between what a system is capable of doing and what it subjectively experiences. Intelligence allows systems to solve problems, analyze information, and make predictions. But consciousness implies, among other things, the presence of subjective experience—such as the sensation of pain or the perception of color.

But it is tempting for people to interpret research findings and warnings about the risks of using AI models in such a way that we can infer presumed motives from observed behavior.

But all of the above are characteristics unique to human psychology and cognitive processes, not to artificial intelligence. In the case of AI, anthropomorphism comes into play, amplified by the Eliza effect—the psychological tendency of people to attribute human consciousness, understanding, empathy, and intentions to computer programs and AI.

If a company grants autonomous AI models access to important—or even confidential—data, decision-making tools, or channels for interacting with the outside world, this could be dangerous. As Anthropic warns, among other things, this could damage the company’s reputation.

And this could affect the company's market valuation and, in the worst-case scenario, its fundamentals and balance sheet.

This also raises the following key questions:

— Should you invest in a company that isn't sufficiently protected against "unethical" behavior by AI models?

— Doesn't AI have too much important and confidential information for us to suffer significant harm from its "unethical" use by AI?

This article was AI-translated and verified by a human editor