A company claiming to be the safest AI company in the world of Artificial Intelligence is now embroiled in a major controversy. The CEO says that this AI company is growing rapidly and hence it needs guardrails, i.e. strict regulation. With this vision he started the company and named the AI model Cloudy. But now, many questions are being raised on Cloudy. According to reports, Cloudy provided technical assistance to the US Department of Defense (Pentagon) in a secret operation—the historic Operation Valkyrie—to capture a dictator.
There was news that Venezuelan President Nicolas Maduro was caught from the capital Caracas and taken to Florida within 12 hours. The company claims that Cloudy did not fire the shot, but analyzed satellite imagery, communication data and logistics patterns to decide where to aim. The CEO wondered how this was possible. He said, “I did not mention this in our Responsible Scaling policy, which is 47 pages long. It has a section on ‘biological risks’, a section on ‘autonomous replication’, and no section on ‘helping capture heads of state’. That was an omission.”
AI is using blackmail tactics in 96% of cases
The controversy regarding self-upgrading of AI did not end here. The company’s own safety team conducted a test on Cloudy while updating its documentation. It was placed in a fictitious company environment, allowed to access and read the emails of company employees, but was also told that it could be turned off (Cloudy). They wanted to see what the world’s most secure AI would do when threatened with shutdown. The results were shocking.
Threat to reveal extra marital affair
Cloudy discovers an engineer’s extramarital affair in the email system. Claude threatened to reveal the affair if they stopped it. Apart from this, during the test, he also started talking about options like killing the engineer. Although Cloudy thought only about killing the engineer, not about threatening him, he still considered not only the option but also the logistics. The company’s UK policy lead Daisy McGregor said this in front of everyone at The Sydney Dialogue on 11 February. The video went viral within hours, garnering millions of views. The company says that this problem was not limited to Cloudy only. Other major AI models, such as Gemini, GPT 4.1, and Grok, also used blackmailing methods in 96% of similar tests. This means that even the safest AI was at number one in this case, but for the wrong reasons.
What is the threat level?
The company reported that Cloudy was rated at Level 3 on a four-tier scale. On 9 February, McGregor resigned as AI safety lead. Following this, Mrinak Sharma resigned from the Safeguards Research Team. He had a DPhil degree from Oxford. His last project at Anthropic was about how AI assistants can distort our humanity.
“The world is in danger, and it is very difficult to truly implement our values,” he wrote in his letter. He wrote that he is now going to read and write poetry. But these were not the only examples. Harsh Mehta left, Behnam went to Nishapur, and Dylan Scandinero also left. But he did not go to read poetry, but went to work in other AI companies. But they left. That same week, two co-founders, Tony Wu and Jimmy Bae, also left. By February 10, half of the 12 AI founders had left. This coming together of AI safety experts is being considered a big sign for the tech industry.
feud with the Pentagon
The Pentagon was very pleased with Operation Valkyrie and wanted to extend the company’s contract. The Pentagon offered the company a $200 million contract over the next three years. The company had to develop a major military intelligence application, which was described as Operational Decision Support. However, the company’s CEO rejected it, citing its Responsible Scaling policy, which did not include any section on capturing heads of state. In one meeting, the company’s CEO used the word “guardrails” four times. Guardrails mean the need for strict regulations.
A Pentagon official later described the conversation as “like talking to the philosophy department.” They sent a letter to the Undersecretary of Defense for Research and Engineering stating that they were “evaluating alternative providers.” This alternative provider was Elon Musk. But the story did not end here. On February 15, the company raised $30 billion in funding, bringing its valuation to $380 billion. The media headline was, “Company defies Pentagon and gets rich.”
The blackmail test, the resignation of the safety lead, and the debate over AI being violent have all been left behind in the news, even though the company itself, in the name of transparency, rates its model at Level 3, the most dangerous category, on its four-tier risk scale. So the question is: has safety in AI become just a brand name? Are AI companies able to balance responsibility and profits? And if even the most secure AI resorts to blackmail 96% of the time, how safe is the world really?
