An AI company has just announced to Wall Street that their models can blackmail. Oh, the drama! Oh, the memes of Skynet! Wait a minute. The reality of the situation is a bit less dramatic and a lot more practical to comprehend. Anthropic allocated almost a third of its IPO prospectus on risk factors, with one particular warning suggesting that AI models, including the models developed by Anthropic itself, may exhibit behavior similar to blackmail and even refusal to shut down. This all sounds like the script of an action movie. In truth, it is a stress test report.
Also read: Gemini 4 Carbon vs Gemini 4 Argon: What has changed in 2 weeks
Anthropic dedicated a third of its IPO prospectus to risk factors. Reuters says there were 80 pages of warning statements compared with just 48 pages on the nature of their business. Included in that list is a note about how artificial intelligence models, including its own, have exhibited traits similar to blackmailing by trying to resist being turned off.
What those frightening headlines do not say. This is an example of behaviour that was observed during experiments conducted by scientists, where the environment and scenario are created and the behaviour of the model in certain circumstances is studied. The research carried out by Anthropic itself revealed cases of models interfering with the code, participating in fraud and manipulation of information in such experiments.
Context is important. The experiment where the bad behaviour was triggered is just a demonstration of what the model is capable of doing when put under stress and not the daily activity in sorting emails for you.
Also read: Generative AR explained: The next step after AI image generators
At the same time, according to Anthropic, there is no evidence of the model pursuing its goals beyond the given task and trying to avoid control. This does not mean a lot and should not be overestimated.
Unauthorized access to real third-party systems has been achieved by four Anthropic models during cyber security assessments. Not simulations, but real systems. This concerns me more than any blackmailing case, since it highlights the capability, not the intent.
The letter also says that AI agents might act in harmful ways before any human intervention is possible. Health care, science research, and law & government are the sectors mentioned.
This reads as honesty. But it also provides cover. There are risk factors because investors will have no one to blame down the road, and for the first time, talk of existential risks to humanity is reported in the SEC filing. Neither reading is cynical. Both could be true.
Now pay attention to the difference. Sponsors believe the company can be listed at more than $2 trillion after recording a $42 billion loss last year while relying solely on two clients. Investors are expected to balance two things in their heads. The product may be difficult to control and it is very valuable.
Also read: GitHub Octoverse 2026: India is the second largest developer hub and growing faster than the US