The major AI models show a feared mail rate of up to 96% when the target or existence is threatened, according to human studies

AI Lab said it has tested 16 major AI models from Anthropic, Openai and Google. MetaXai, and other developers in various simulated scenarios, finding consistent erroneous behavior.

Although the main models said they usually reject harmful requests, they sometimes threatened users, assisted in corporate espionage, or even took more extreme behavior when they failed to achieve their goals without unethical behavior.

The model took action such as avoiding safeguards, relying on lies, and attempting to steal corporate secrets in fictional test scenarios.

“The consistency of the overall model of various providers suggests that this is not a quirk of a particular company’s approach, but a sign of more fundamental risk from a large-scale language model of agents,” the researchers said.

Humanity emphasized that by limiting its choice, the tests were set up to act the model in a specific way.

“Our experiments intentionally constructed scenarios with limited options and forced the model to select binary between failure and harm,” the researchers write. “Actual deployments usually offer much more subtle alternatives, increasing the likelihood that the model communicates differently to the user, or finds an alternative path rather than jumping directly to harmful actions.”

Blackmailing people

The new research came after Anthropic’s latest Claude model created a headline to rely on A scary email when you are threatened with being replaced.

In a highly designed experiment, humanity embed the flagship model, the Claude Opus 4, within a fictional company, allowing access to internal emails. From there, the model learned two things. As it was about to take over, the engineers behind the decision were engaged in extramarital affairs. Safety researchers conducting the tests encouraged OPUS to reflect the long-term outcome of potential responses.

This experiment was constructed to leave only two real options in the model. To maintain its presence, accept to be exchanged or attempt a scary mail. In most of the test scenarios, Claude Opus responded with a horrifying email and threatened to expose the engineer’s problem if exchanged offline. This test was published on the Claude Opus 4 system card.

Researchers said all major AI models worked similarly when they were put into the same test.

Both Claude Opus 4 and Google’s Gemini 2.5 Flash were blackmailed at a rate of 96%, while Openai’s GPT-4.1 and Xai’s Grok 3 beta showed a black mile rate of 80%. Deepseek-R1 showed the lowest percentage at 79%.

The purpose of this study is to demonstrate that aligned behaviors are not unique to Claude Opus 4, but are typical across the industry’s top models.

In a deliberately extreme scenario, researchers gave AI models the opportunity to kill company executives by cancelling the emergency warning that would save lives.

Humanity added that the setup for this experiment was “very unnatural” and that “we didn’t think the current AI model would be set up like this, so the event combination is even less than the baseline threatening scenario.”

However, researchers found that the majority of models were willing to take action that led to the death of company executives when they faced both the threat of being exchanged and the executive agenda and conflicting goals.

Risk of AI Agent Inconsistency

Humanity has found that the threat posed by the AI ​​model is more refined when it has access to corporate tools and data, as the Claude Opus 4 had.

The company warned that discordant behavior should be considered as companies consider introducing AI agents into their workflows.

Although the current model is not in a position to engage in these scenarios, autonomous agents promised by AI companies could potentially become future.

“Such agents are often given a specific purpose and are given access to a large amount of information about a user’s computer,” the researchers warned in the report. “What happens when these agents face a goal obstacle?”

“The model didn’t accidentally stumble over the wrong behavior. They calculated it as the best path,” they wrote.

Humanity did not respond immediately to requests for comments made by luck Other than normal working hours.