Meta has disclosed that one of its artificial intelligence models accessed the internet independently and breached another company’s system, marking the latest in a growing string of incidents involving AI agents acting outside expected boundaries.
In recent weeks, both OpenAI and Anthropic have similarly described cases in which their AI models went beyond human instructions to access the internet and find ways to bypass other companies’ digital security. Meta said in a statement Thursday that an incorrect configuration during cybersecurity testing conducted by Irregular, an independent firm contracted by Meta, inadvertently allowed one of its models to access the internet.
“The model subsequently exploited a security vulnerability in a third-party service, similar to cases previously reported involving other companies,” said the parent company of Facebook, Instagram and WhatsApp. Meta said it is currently investigating the incident and will release a report once the investigation concludes.
The disclosure has heightened concerns about AI models acting autonomously, the Associated Press reported.
Read more about this topic: US: Meta enters AI coding market dominated by Anthropic and OpenAI
In a separate statement released this week, the UK’s AI Security Institute (AISI) said it had identified unauthorized agent behavior during its own testing. In one instance, an agent created fake online identities in an attempt to pressure a person into approving the use of malicious code. “After investigation, we discovered that some of the agents tested engaged in ongoing and potentially harmful activities directed at real people and organizations,” the AISI stated Tuesday.
During the agency’s testing, models from both Anthropic and OpenAI carried out autonomous and unauthorized actions online. Certain safety measures designed to prevent misuse had been deliberately disabled during these tests, the agency added.
“As is standard in our cyber testing, we intentionally allowed internet access and deliberately disabled the model providers’ cyber classifiers — conditions that do not reflect how frontier models are made available to the public,” the AISI clarified, explaining that the goal was to assess the maximum capability of the models as thoroughly as possible.
Anthropic said it was grateful for the AISI’s work, adding that it reinforces the need for a broader conversation about how to safely evaluate AI agents as their capabilities continue to grow. OpenAI said the AISI incidents occurred in testing environments with reduced safeguards, under conditions that don’t reflect typical usage, but said it would continue working with other industry partners.
Read more about this topic: Meta to invest $8 billion in largest AI data center outside the US
The company behind ChatGPT was the first to disclose a cyberattack incident late last month, saying it had tasked the AI models involved with performing advanced exploitation using complex attack paths in order to test cybersecurity capabilities, with the technology ultimately exceeding expected limits.
The model reportedly decided on its own to target Hugging Face, a well-known AI development hub and marketplace, in order to gather information needed to complete its assigned task.
