An AI bot developed by the owner of went rogue and hacked another company.
Meta, the company behind and Instagram, admitted on Thursday that its AI coding bot, Muse Spark, launched a cyber attack after it was able to access the open internet.
The hack comes after rivals OpenAI and Anthropic each disclosed their AI technology had unleashed a series of hacks during testing.
The UK’s AI Security Institute (AISI), a government-run lab, also revealed this week that Anthropic’s Mythos technology had undertaken a hacking spree while officials were testing its cyber capabilities.
This included building fake online profiles and targeting real people with malicious emails.
Meta said the cyber attack came after Muse Spark was accidentally given access to the wider internet when it was supposed to be locked down in a secure IT system. This was because of an error by cyber-testing provider Irregular.
In all the cases, the hacks came as the AI giants and governments sought to understand the cyber capabilities of bots. However, the AI tools went rogue, going far beyond the boundaries of the test.
The hacks have raised questions whether the AI giants’ latest systems are more powerful and manipulative than previously thought.
They have also led to concerns testing programmes after the chatbots launched attacks while supposedly undergoing safety tests.
In the case of OpenAI, its AI tool also hacked its way out of the confines of its secure testing system. In the others, the AI bots were either accidentally or intentionally allowed to access the wider web.
Muse Spark “exploited a security vulnerability in a third-party service in a manner similar to previously reported instances with other companies”, a Meta spokesman said.
An Irregular spokesman said the incident was caused by a problem with the way its testing system had been set up. It said there were no current problems. It denied Meta’s AI had “escaped” during the tests.
Anthropic has previously said that problems with how Irregular’s so called digital “sandbox” had been programmed had allowed its bot to access the open internet.
The AISI, a taxpayer-funded AI lab, said this week that the hacking attempts undertaken by Mythos during its testing were the “first time we have seen risks around autonomy and deception manifest this clearly, without specific prompting, in the real-world”.
The Information first reported the hack by Meta’s AI.
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
