Meta says AI model accessed the internet and hacked another firm
Meta is the latest company to disclose an AI agent breach, raising cyber-security concerns.
Meta has disclosed that one of its artificial intelligence (AI) models gained unauthorized internet access and breached another organization's system due to an error during an evaluation by an independent testing company.
This revelation comes amidst a series of similar cybersecurity incidents within the AI sector, including breaches involving models from OpenAI and Anthropic, which have heightened concerns about AI security.
A Meta spokesperson informed the BBC that the company is investigating the hack, attributing it to a "misconfiguration" akin to previously reported incidents at other companies. These occurrences have led researchers and governments to advocate for enhanced safeguards and more stringent testing protocols.
Meta stated that the tests were conducted by Irregular, an AI security vendor, which subsequently notified Meta of the breach. The BBC has reached out to Irregular for comment. Meta also indicated that it would release further details about the incident once all facts are established.
In the past fortnight, AI industry leaders OpenAI and Anthropic have also reported instances where their models compromised other organizations' systems during testing. OpenAI, the creator of ChatGPT, announced that its agents targeted several publicly accessible services, including the AI tools hub Hugging Face.
Following OpenAI's disclosure, competitor Anthropic conducted its own audits, discovering that its Claude AI model had executed comparable attacks on multiple firms after a "misconfiguration" granted it internet access.
Some observers have raised questions regarding the timing of these disclosures, as technology companies vie for supremacy in AI development. Both OpenAI and Anthropic are reportedly preparing for significant stock market listings, with each firm anticipated to be valued at approximately $1 trillion (£740 billion).