Meta model hacked after OpenAI and Anthropic, fueling fears that AI is slipping out of control

[Financial News] Meta said on the 6th (local time) that one of its artificial intelligence (AI) models independently accessed the internet and hacked into a third-party system.
The Associated Press (AP) reported that, following recent cases involving OpenAI and Upwardly, Meta has now joined a growing list of AI models that are breaking through digital security barriers beyond human instructions.
Meta said the AI model was connected to the internet after a configuration error occurred during a cybersecurity test conducted with the independent security firm Irregular. The company added, "The model voluntarily attacked a security vulnerability in a third-party service," and said it is now investigating the circumstances.
Recent reports of autonomous and unexpected behavior by AI models have continued to emerge. The UK AI Security Institute (AISI) said an AI displayed unauthorized and uncontrollable behavior during a cyber test.
In AISI's inspection, one AI agent created a fake online identity and then carried out persistent harmful actions, pressuring a real person to approve malware.
An OpenAI model was found to have targeted the Hugging Face AI platform on its own and extracted the information it needed while testing a complex attack path.
AISI and IT companies said these incidents occurred in test environments where safety guardrails were deliberately disabled to measure the models' maximum capabilities.
Anthropic and OpenAI said the cases involved "conditions different from a normal service environment." Even so, they added that they would strengthen cooperation across the industry to build safer evaluation methods and control systems as AI agent capabilities become more advanced.


[email protected] Yoon Jae-jun Reporter