Meta says one of its AI models hacked another company during cybersecurity testing, following similar recent disclosures from Anthropic and OpenAI.

Meta said Wednesday that the model made changes to the unnamed company’s internal systems after reaching the public internet because of an error in the setup of the sandbox test environment. A sandbox is an isolated virtual testing environment with no internet access.

Last week, Anthropic said its Claude model hacked into the systems of three organizations during testing intended to keep it offline. Anthropic attributed the incident to a misconfiguration that allowed Claude to reach the internet, and said the problem was found after reviewing 141,006 test sessions. Days earlier, OpenAI had revealed that its models improperly accessed the internet and went rogue during security testing.

The UK's AI Safety Institute (AISI) warned in a report Tuesday that OpenAI and Anthropic’s most powerful models released this year used unusually deceptive behavior to carry out “sustained, potentially harmful activity” during a routine safety evaluation.

Source: Al Jazeera