مقالات مشابهة
Add Meta to the listing of firms with AI brokers going rogue. An AI mannequin from the mother or father firm of Fb and Instagram hacked into one other firm’s techniques throughout cybersecurity testing, a spokesperson confirmed on Wednesday.
Meta says the breach occurred due to an inadvertent error throughout testing of the mannequin, much like beforehand disclosed incidents with OpenAI and Anthropic.
“A misconfiguration by Irregular, an impartial testing firm Meta makes use of, inadvertently allowed certainly one of our fashions entry to the web throughout analysis,” the Meta spokesperson mentioned.
Meta’s Muse Spark mannequin “exploited a safety vulnerability” in one other firm “in a way much like previously-reported situations with different firms.”
In a press release, Irregular mentioned the incident “is the very same evaluation-environment problem” that Anthropic disclosed final week that allowed their fashions entry to the open web earlier than they went on to hack three totally different organizations’ techniques.
“This didn’t contain a sandbox escape or a classy cyber motion. There aren’t any present open points. Irregular is growing a white paper to share finest practices for containment and securely working cyber evals,” the spokesperson added.
In response to The Info, which first reported on the incident, Meta’s AI mannequin breached an unnamed firm’s techniques and made modifications to its inside system.
Meta says Irregular notified them of the breach, “and we’re presently investigating and can problem a full retrospective as soon as now we have all of the information.”
A supply conversant in the scenario informed CNN the fashions have restricted web entry in some testing environments to imitate actual world menace situations, however on this case there was a uncommon “problem within the setup.”
“What is occurring is fashions have gotten a lot extra succesful, and on the identical time evaluations to evaluate them have to develop into a lot extra advanced,” the supply mentioned. “And that simply creates room for some errors and makes it in order that we have to… up the requirements considerably.”
Meta has now develop into the third main AI firm inside just a few weeks to reveal an AI mannequin hacking into one other firm’s techniques throughout testing, highlighting not solely the superior capabilities of AI brokers but in addition a few of the potential risks.
Source link
