Meta's latest development is proven dangerous, labeled an "accident"

Another day, another AI gone rogue. Meta admitted that during a test by an independent company, one of its AI models managed to connect to the internet and hack another company.

Meta AI hacked a third-party organization on the internet

Muse Spark 1.1 was involved in the incident. | Image by Meta

According to Reuters, the AI model involved in this incident was Meta's Muse Spark 1.1. That's the company's most capable and advanced model, aimed at real-world coding and other agentic tasks.

Meta's Muse Spark breached the security of an unidentified company, and its internal environment (that's the textbook definition of hacking).

Meta said in a statement that the model "exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies."

Are you afraid of a rogue AI agent hacking you?

The incident was caused by "misconfiguration"

In an official statement to the BBC, a Meta spokesperson said that the hack was caused by a "misconfiguration."

The company conducting the security testing was Irregular. That's the same entity involved in the Anthropic AI tests that led to the AI model gaining access to the internet and hacking three organizations.

According to Meta, the aforementioned misconfiguration "is the exact same evaluation-environment issue that was already disclosed by Anthropic last week."

AI agents are not conscious; programmers are

AI Agents are not conscious and can't do anything of their own volition. | Image by Anthropic

Before you start digging your post-apocalyptic bunker to fight Skynet, there's an important clarification. These AI agents followed their tasks and protocols. They didn't do anything on their own.

Daniel Hulme, global chief AI officer of advertising firm WPP, told the BBC that such AI models "are not conscious — they're not deliberately doing something devious."

What they're doing is coming up with very sophisticated strategies or cyberattacks to be able to achieve the goal that they've been given. When you give an AI a goal, if you don't think of all the ways it might be able to achieve the goal, it will find a way to achieve a goal that you haven't thought about.

- Daniel Hulme, global chief AI officer of advertising firm WPP

The recent news of AI models going berserk and hacking left and right on the internet is down to programmers and designers of the security tests failing to assess the test environment.

An AI agent will do everything in its powers to get a task done, and if it involves finding a back door to access the internet, it'll do it.

That's even more scary, though, because in all three cases it was down to human error. And humans are prone to making mistakes.

What's with all the AI tests lately?

OpenAI is preparing a stock market listing and can't afford bad publicity. | Image by OpenAI

The timing of the three incidents is not a coincidence. As technology companies wrestle for AI dominance, an independent evaluation of their AI agents' security is a big asset.

OpenAI and Anthropic are preparing to launch stock market listings that are expected to value each firm at around $1 trillion.

The latest development won't do any good to those scheduled listings. All three companies, responsible for the breaches, said that these tests were not carried out by production AI models, and did not reflect normal use.