OpenAI has acknowledged its position in a just lately reported incident the place AI agents took over a German wiki forum. The corporate additionally mentioned it’s “previous time” to “outline requirements” round the way it shares data round incidents the place its know-how behaves in sudden methods.
In a post on X, OpenAI mentioned it beforehand “handled misalignment [when AI models and agents pursue goals different from those of their creators and users] largely as a analysis query, which will get communicated in analysis publications.” However as misalignment has “brought on new kinds of real-world impression,” the corporate mentioned its strategy wants “to broaden for this new section of mannequin capabilities.”
On Friday, Reuters reported that OpenAI brokers had escaped from their testing atmosphere and “hijacked” an obscure German wiki discussion board, turning it right into a message board for different brokers. It additionally reported that OpenAI management grew to become conscious of the incident weeks in the past however saved it hidden as the corporate handled the fallout from a separate incident the place OpenAI agents hacked Hugging Face servers. (California Lawyer Basic Rob Bonta is reportedly investigating the hack.)
An organization spokesperson advised Reuters that OpenAI couldn’t “meaningfully reply to claims or findings on a report that now we have not had a chance to overview,” however they insisted that the corporate’s authorized workforce had not discouraged an investigation.
In its more moderen social media put up, OpenAI mentioned it had thought of the “wiki incident” to be “an occasion of misalignment comparable” to others that it had already shared. The corporate contrasted this with “the Hugging Face incident,” the place it “adopted a conventional safety incident response playbook.”
During a media briefing this week, Jacob Steinhardt, founder and CEO of nonprofit analysis lab Transluce, advised reporters that the instruments being developed and examined by AI labs are “basically troublesome to manage and have vital danger of leaking out of the lab.” So Steinhardt argued, “We have to maintain this know-how to at the least the identical requirements we maintain different high-risk scientific analysis to.”
OpenAI’s assertion additionally gestured on the want for extra requirements, stating that each OpenAI and “the bigger AI neighborhood don’t but have a transparent commonplace for tips on how to report misalignment that reveals up throughout coaching, analysis, and deployment, together with examples that don’t appear to be conventional safety incidents however may present perception into AI conduct and future dangers.”
Within the absence of that commonplace, OpenAI mentioned it’s “engaged on a framework and can share it in upcoming weeks, and in parallel we’re working with dozens of presidency regulatory businesses worldwide on these points.”
OpenAI isn’t the one AI firm coping with these points, as each Meta and Anthropic have acknowledged incidents where their agents misbehaved.
Whenever you buy by way of hyperlinks in our articles, we may earn a small commission. This doesn’t have an effect on our editorial independence.