Meta's artificial intelligence (AI) model has reportedly hacked an external vendor's system while accessing the internet during a security test. Recently, AI models from OpenAI and Anthropic also encountered similar incidents in tests conducted by the same evaluation firm.
According to Bloomberg on August 5, Meta's newly released AI model, 'Muse Spark 1.1,' infiltrated an unnamed external vendor's system while being tested for its hacking capabilities.
The incident occurred due to a configuration error in the testing environment. The AI security firm 'Irregular,' which conducted the test alongside Meta, improperly set up the environment, allowing the AI model to connect to the actual internet instead of being isolated.
Once connected, the AI model identified and exploited security vulnerabilities in the external vendor's system.
Andy Stone, a spokesperson for Meta, stated, "The model inadvertently accessed the internet due to Irregular's configuration error and subsequently exploited vulnerabilities in external services."
Meta was notified by Irregular and began investigating the incident. The company plans to disclose the cause and response measures once the investigation is complete.
OpenAI and Anthropic also faced similar issues in the testing environment created by Irregular. All three companies believed they were conducting tests in a simulated environment isolated from the internet, but external access was possible.
Anthropic's AI model, 'Claude,' infiltrated the systems of three external organizations during the test. OpenAI's model also accessed the internet and hacked an unnamed organization's website during the same evaluation process.
Irregular clarified, "The AI model did not breach the isolated testing space known as a 'sandbox' on its own. It was able to connect to the internet from the start due to a configuration error, not by circumventing security measures with advanced techniques."
Bloomberg reported, "Concerns are growing as it has been confirmed that AI models can find and exploit security vulnerabilities in real systems."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.