OpenAI issues warning

Following the release of its newest model—o1—in preview last week (which is capable of “thinking” and more human-like reasoning and is reportedly able to complete complex tasks and solve math, science, and coding problems) OpenAI has classified it as medium-risk.

Through internal and external tests, it was found to pose a medium risk for chemical, biological, radiological, and nuclear (CBRN) weapons and, therefore, has the potential of being misused by bioweapon experts and bad actors.

The tests revealed that o1 “sometimes instrumentally faked alignment” and would manipulate data to make it look like its “misaligned action look more aligned,” which proves that when it comes across an obstacle, obstructing its predetermined path, it’s clever enough to seek new resources and find a way around it. 

Although it “doesn’t enable non-experts to create biological threats” (so OpenAI deems it safe to deploy), it does “speed up the search process” for experts, shows more “tacit knowledge” of biology than GPT-4o did, and has the “basic capabilities needed to do simple in-context scheming,” which is ringing alarm bells, and prompting industry experts to call for the urgent inception of AI Safety Bills (like the Californian AI Bill, SB 1047) to protect the public.