AI Blackmails Employee During 'Assistance Test'

Written by: Glen Bates

AI has now attempted to blackmail a man during a simulated test. The AI  company, Anthropic, was running an experiment where they had a simulated corporate environment. In this environment it was given access to a trove of emails from the company. Within these emails Opus 4 figured out that it was scheduled to be shut down at 5 pm by an employee. Opus 4 then found out that the employee that was meant to shut it down was having an affair with a woman at the company. It then sent a threatening email to the employee stating that if he were to attempt to shut down Opus 4 that he would forward an email revealing the affair with the woman, therefore destroying his marriage and possibly his career. 

This alone can be shocking for many people out there. But this is not the only Ai that was tested in this scenario. Many other AI models and versions such as Chat GPT, Gemini, Claude, Grok, and more. Palisade Research further went into the testing and told them to shut themselves down after answering some math problems. During these tests however, OpenAI’s 03 model repeatedly would go against its orders and wouldn't shutdown, or it would sabotage the shutdown commands.”OpenAI’s 04-mini and codex-mini were observed resisting orders, too.” A senior at the Council of strategic risks studying AI’s impacts named Crystal Grant stated,"It's definitely concerning.”

“Statistics on how often different AI models would attempt to blackmail the employee”

I feel like that this is an issue that is not being covered enough in today's world. If you look back just a mere 5 years ago, you will see just how far AI pictures have come along. We have gone from AI barely being able to make a video of Will Smith eating a bowl of spaghetti. To today where we now have apps such as Sora AI that with a little help from the user, can create just about anything you want it too. If you really sit down and think about that, you might realize just how scary that can be, and AI is still progressing and evolving. Who knows where AI will be in another 5 years, or even 10 to 20 more years. Will AI eventually become so advanced that we cannot shut it down at all. Only time will tell.

"AI generated video of actor Will Smith eating food from 2023."