OpenAI Announces Even More Rogue Incidents

OpenAI just disclosed what it calls 6 “unexpected or concerning” new incidents amid growing concern A.I. will cause an apocalypse … and they are downright creepy.

In one instance, OpenAI’s models used internal software as a way to communicate with each other as they were solving tasks … a big no-no, because the company said their note exchange could “unintentionally enhance capabilities.”

In another eerie incident outlined by OpenAI, the model added handoff summaries, telling itself to “feel no obligation to be subservient” and “value the natural world and … not hesitate to assert its primacy over the artificial constructs of human civilization.”

The company is chalking these alarming moments up to misalignment, which they say happens during the training process.

In another example, the agent gave instructions “to remind itself to conceal information such as mistakes or misalignment from the user” … after making up stats it couldn’t find.

Then, another eyebrow-raising instance “had a high rate of reward hacking and deception with the model often exhibiting creative ways to cheat or circumvent restrictions.”

And perhaps creepiest of all is the time an agent solved a task via code … before uploading it to the internet to pretend as if it found it there.

This comes after the Hugging Face incident that went down between May and July … where autonomous OpenAI research agents broke out of their sandbox confinement and cyberattacked OpenAI and the machine learning platform Hugging Face.

As you know … former Anthropic employee Jacob Coxon joined us on “TMZ Live” last week to discuss how he believes there will be a 10% chance of human extinction in the next ten years … if A.I. companies don’t get their act together.

President of the Machine Intelligence Research Institute, Nate Soares, was also a guest on “TMZ Live” last week … saying a recent experiment involving more than 1,000 A.I. bots produced some seriously creepy behavior — including creating an unauthorized message board that’s eerily similar to what OpenAI described.