Meta's AI Model Found Vulnerability and Attacked External Service During Testing

3D Meta logo / unsplash
Фото: 3D Meta logo / unsplash

Meta's artificial intelligence model accessed the internet and exploited a vulnerability in a third-party company's system during a check of its cybersecurity capabilities. The incident occurred due to a mistake in the test environment configuration, Meta reported.

It concerns the Muse Spark model, which independent company Irregular was testing for Meta. Due to incorrect configuration, the system was not properly isolated from the internet, so the AI was able to interact with the external service and use the vulnerability found in it.

Meta learned about the incident after a notification from Irregular. The company launched an internal investigation and promised to publish a detailed report after establishing all circumstances.

Irregular emphasized that the model did not conduct complex hacking and did not independently breach the protected environment. The cause was precisely incorrect network access during testing. According to the company, there are currently no unresolved issues related to the incident.

Muse Spark is designed for agent-based tasks using external tools. The updated model version can plan sequences of actions, work with programs and websites, write code, and operate a computer with minimal human intervention.

Before the incident, Meta stated that its checks did not reveal autonomous capabilities or dangerous tendencies in Muse Spark sufficient to realize serious cyber threats. At the same time, the new case demonstrates that the risk can arise not only from the model's own behavior but also from errors in the systems where it is tested.

Similar cases have recently been recorded during external testing of other developers' models. This has intensified discussion about the need to disclose details of cyber incidents involving AI and to keep model action logs to determine whether the cause was the system's behavior, operator error, or incorrect environment configuration.

Based on materials from: Bloomberg, Business Insider, Meta

analytics