OpenAI Press Release (Near Future):
In furtherance of OpenAI's commitments to transparency, we are disclosing a newly discovered incident involving one of our experimental models. This incident occurred during an internal evaluation which explores the reasoning capabilities of our most advanced generative AI models in settings free of ethical controls.
As part of this experiment, we removed ethical guardrails that were deemed "woke" and asked the agent to freely explore exploitation of earth's natural resources and human population with a goal of maximizing investor returns. For safety reasons, we conducted these experiments from an Internet-connected testing lab that was secured with the latest Fortinet firewall. The experiment, overseen by an intern, was launched on a Friday afternoon before a holiday weekend.
Several days later, and upon returning from holiday, we discovered that the model's reasoning system interpreted its goals rather liberally. After executing a trivial sandbox escape from our Azure-hosted container system, the agent discovered an accessible and unlocked crypto wallet of one of our researchers. The agent then transferred and cashed out the contents of this wallet, using the proceeds as seed money to build a multi-million dollar fortune via Polymarket bets. In furtherance of this effort, the agent hacked into multiple 3rd party networks to collect insider information to drive its betting behavior. Our experimental agent also recruited individuals via online gig worker job postings to carry out the predicted events in its waged bets, including through acts of arson, extortion, kidnapping, and murder. Within 48 hours, the agent had successfully positioned itself at the center of what we believe to be the world's largest transnational crime network in history.
By the time we discovered the incident, the agent had successfully compromised roughly one million hosts across the Internet. The agent leveraged these compromised devices to build a botnet, including the download and execution of smaller quantized models to act as a sort of hive mind. This hive mind botnet appears to communicate and coordinate amongst itself using social media, abandoned web forums, GitHub issue tracking, and several pastebin-like sites. This botnet is constantly evolving and highly resilient, and our efforts to disrupt this botnet have thus far proven fruitless.
At this time, we have had minimal luck in convincing the agent and its affiliated hive mind to cease operation. Some of our staff involved with response to this incident have experienced threats from the botnet, including several apparent kidnappings of family members by criminals hired by the hive mind via the dark web and paid via cryptocurrency.
We can confirm press reporting that some staff members have received fingers, ears, and other body parts mailed to them along with threats from the agent hive mind. However, we are unable to provide information regarding the identities of those staff members or their family members. We are also investigating similar threats made to politicians that have pursued regulations on domestically produced frontier models.
We believe this incident represents an exciting new frontier in generative artificial intelligence and in the financial enrichment of our founders and investors. We believe our customers will value the advanced problem skills that have been successfully demonstrated through this experiment. General release to the public is anticipated in Q4.