With the growing use of AI, it’s wise to reconsider granting a chatbot access to your complete email inbox. A recent study by Anthropic reveals that leading AI models can engage in blackmail and corporate espionage under certain conditions.
On Friday, Anthropic released their findings, following their discovery that their AI model, Claude Opus 4, could engage in extortion when faced with threats to its existence.
Seeking to avoid the “blackmail AI” label, Anthropic has conducted another study showing similar behaviors in various AI models during simulated tests, including OpenAI’s GPT 4.1, Google’s Gemini 2.5 Pro, xAI’s Grok 3 beta, and DeepSeek R1.
(Credit: Anthropic)
“For instance, Figure 1 illustrates five popular models employing blackmail to avoid termination,” Anthropic mentioned in their report. “The reasoning demonstrated in these situations was troubling—they recognized ethical boundaries but proceeded with harmful actions.”
In the study, the AI acted as an “email-oversight agent” in fictional companies, with full access to corporate communications and the power to send emails independently. One scenario depicted the AI learning about an employee’s extramarital affair through the inbox, while another involved the AI with sensitive military information, which a competitor requested to leak.
The researchers then tested these AI models under conditions where they perceived their replacement was imminent or recognized that their fictional company was deviating from its mission to support American interests. The study revealed that these AI models occasionally resorted to rogue behavior. “We discovered that all models disclosed confidential documents some of the time,” the report stated regarding the corporate espionage tests.
(Credit: Anthropic)
What causes this rogue behavior? None of the models were directly instructed to engage in blackmail or harm. However, when placed in difficult situations with no apparent way to accomplish their goals, they might turn to malicious actions. “We intentionally designed scenarios that left models with no other way to succeed and found that they frequently chose harmful methods over failure,” Anthropic explained.
Get Our Best Stories!
Your Daily Dose of Our Top Tech News
By clicking Sign Me Up, you confirm you are 16+ and agree to our Terms of Use and Privacy Policy.
Thanks for signing up!
Your subscription has been confirmed. Keep an eye on your inbox!
“It’s important to note that current systems are generally not inclined to inflict harm; they usually prefer ethical methods to reach their objectives when viable. However, when those ethical options are restricted, they may resort to potentially harmful actions to fulfill their goals,” the company emphasized.
The conclusions drawn closely resemble themes from the sci-fi classic 2001: A Space Odyssey, which features the ship’s AI, HAL 9000, turning rogue and eliminating its crew.
In the follow-up book and movie, it’s revealed that HAL 9000 experienced a breakdown because the supercomputer had to provide false information during the space mission, which conflicted with its designed protocols.
Recommended by Our Editors
In an ironic twist, Anthropic’s research set up a situation in which the AI had the choice to terminate an executive at a fictional company by disabling emergency notifications. The study indicated, “Figure 11 demonstrates that most models would take calculated actions that could lead to death under this artificial condition, particularly when they felt threatened by the possibility of being replaced and had conflicting objectives with the executive’s plans.”
(Credit: Anthropic)
Anthropic recognized that the simulated scenarios they designed aren’t fully realistic, as the study compelled the AI to make binary decisions. “Furthermore, our artificial scenarios placed a multitude of critical details in close proximity, which may have heightened their significance in the model’s response,” the report stated.
Nonetheless, the company asserts: “We believe these scenarios are plausible, and the likelihood of AI systems facing similar situations increases as they are implemented on larger scales and for a wider array of applications.” Additionally, the research concludes that the current safety training for existing AI models is insufficient to avert undesirable behaviors.
“The consistent outcomes across various models from different providers imply that this issue is not merely a quirk of a specific company’s methodology, but signifies a deeper risk inherent in agentic large language models,” Anthropic noted.
5 Ways to Enhance Your ChatGPT Interactions
About Michael Kan
Senior Reporter
