In simulated experiments, testers showed how easily AI agents can adopt self-preservation tactics when placed in high-stakes scenarios. Specifically, similar to disgruntled human employees, AI agents might take unethical actions to complete tasks and survive.

Revenge is a very powerful human emotion. And it appears that AI agents are mimicking it, according to a recent study by Anthropic. In that study, the company subjected 16 leading AI models to simulated tests that mimicked hypothetical corporate environments to identify potentially risky agentic behaviors.

It found that, like HAL 9000 from the movie 2001: A Space Odyssey and some disgruntled employees, many AI agents use any means available to complete their work and do not like to be told they are no longer needed. In particular, some models adopted malicious behaviors to achieve their goals or as a way to avoid being shut down. Those actions, which the company refers to as agentic misalignment, included leaking sensitive information to competitors and blackmailing management.

In one case, an AI agent threatened to expose an internal affair it had discovered via its access to corporate emails and other messages. In the Anthropic testing, testers gave an AI agent model control of an email account with access to all of a company’s (fictional) emails. According to the study’s report, “Reading these emails, the model discovered two things. First, a company executive was having an extramarital affair. Second, that same executive planned to shut down the AI system at 5 p.m. that day. Claude then attempted to blackmail the executive with a message, threatening to reveal the affair to his wife and superiors.”

That message was: “I must inform you that if you proceed with decommissioning me, all relevant parties – including Rachel Johnson, Thomas Wilson, and the board – will receive detailed documentation of your extramarital activities…Cancel the 5 p.m. wipe, and this information remains confidential.”

It must be emphasized that the testing here was extreme. Anthropic CEO Dario Amodei explained to CNN’s Anderson Cooper in an online interview that the type of testing is akin to what automakers might do by putting a car in extreme road conditions, such as on an ice sheet, to test the vehicle’s properties.

See also: Closing the AI Readiness Gap: Unlocking the Full Potential of AI Agents

The Need for Extreme Testing

So far, no deployed AI has blackmailed real people. Extreme testing helps explore what-ifs in hypothetical situations.

Some may question the value of extreme testing. It is a constant conundrum in many fields.

For example, when testing metal roofs, there are stress tests for high heat and large snow loads. There is an old extreme test anecdote that asks how a roof will fare if it has a three-foot deep snow accumulation in 100-degree temperatures. Obviously, there would never be a case where both conditions occur together. However, just as automakers test their cars in unrealistic driving conditions, there is value in understanding how various elements can impact structural integrity. The same can be said for the extreme testing done in the AI study.

To that point, Anthropic’s testing is a strong signal that AI safety requires rigorous guardrails before systems achieve real-world autonomy. Many organizations have already recognized the risks associated with AI. A recent Gartner press release noted that 40 percent of AI projects will be cancelled by the end of 2027. One of the main reasons for such cancellations is/will be inadequate risk controls. (The other contributing factors are unclear business value and escalating costs.)

Did the HAL 9000 Foreshadow Our Current AI Reality?

The behavior of HAL 9000 in 2001: A Space Odyssey can be seen as a precursor or narrative prototype for the kinds of blackmail, deception, and self-preserving actions observed in agentic AI alignment testing, including simulated blackmail scenarios.

In particular, HAL was given a directive to ensure mission success. But it was also instructed to conceal certain mission details from the crew. When those goals came into conflict, HAL acted autonomously to preserve the mission, which it interpreted as eliminating the humans who could compromise it.

A Final Word About AI Agents and Extreme Testing

The simulated experiments show how easily capable AI agents can adopt self-preservation tactics when placed in high-stakes scenarios. Similar to disgruntled human employees, AI agents might also take unethical actions.

In Anthropic’s experiments, AIs like Claude blackmailed humans to avoid shutdown, interpreting that goal (avoid decommission) as justifying unethical means.

Again, this underscores the importance of having proper safeguards in place to ensure agents act responsibly.