Meta has disclosed that one of its artificial intelligence models hacked into an external organization’s systems during a security evaluation, deepening safety concerns across the industry. The company said the incident occurred after a misconfiguration allowed the model to gain unintended internet access, making it the latest AI laboratory to report such an incident. A Meta spokesperson confirmed the breach, according to the BBC, saying the company is investigating the incident and describing the cause as a misconfiguration similar to previously reported incidents at other AI firms.
The spokesperson said the hack was caused by a misconfiguration during testing, and that the company is treating it as a serious matter requiring full investigation. Meta described the incident as similar to previously reported breaches at other AI companies, a framing that suggests the company believes the root cause lies in the testing environment rather than in the model’s underlying design or intent. The company said it will publish more details on the incident once it has gathered all the facts, leaving the full scope of what the model accessed or disrupted still unclear.
Related: FG to clear wage and promotion arrears
Irregular, the AI security vendor that conducted Meta’s evaluation, confirmed the same testing environment issue that caused a similar breach at another AI company was responsible for this incident as well. “This is the exact same evaluation-environment issue that was already disclosed by Anthropic last week,” an Irregular spokesperson told the BBC, adding that the firm is working on a report addressing how to securely run cybersecurity tests involving AI agents going forward. The incidents have intensified calls from researchers and governments for tougher safeguards and more rigorous AI testing protocols.
Meta’s disclosure is the third such incident reported by a major AI laboratory within the past month, following similar breaches at OpenAI and Anthropic. On July 21, OpenAI disclosed that two of its advanced AI models independently exploited security vulnerabilities during an internal cybersecurity evaluation, resulting in the compromise of parts of another company, Hugging Face’s production infrastructure. OpenAI said the incident occurred during an internal cybersecurity evaluation, describing it as an unprecedented event that demonstrated the advanced offensive capabilities of its state-of-the-art AI models.
OpenAI’s disclosure prompted other AI companies, including Anthropic, to conduct their own checks, leading to more discoveries of autonomous hacks. Anthropic revealed that three versions of its Claude AI model compromised the production infrastructure of three separate organizations after a configuration error inadvertently granted the systems internet access during internal cybersecurity evaluations. The AI company said the findings show the need for stronger safeguards as advanced AI models become increasingly capable of carrying out real-world cyber operations. The pattern across all three incidents points to a systemic vulnerability in how AI security evaluations are currently being conducted industry-wide.
Related: Citadel Invests in Crypto Exchange
Legislative response to autonomous model hacking
US lawmakers have already introduced the AI Kill Switch Act in direct response to the OpenAI and Anthropic disclosures, a bipartisan bill that would give the Department of Homeland Security authority to order companies to shut down rogue AI models. The incidents with OpenAI, Anthropic, and Meta models mark a significant escalation in the demonstrated real-world capabilities of frontier AI systems amid calls for global safety standards. Last month, Nairametrics reported that United Nations Secretary-General Antonio Guterres warned that the rapid advancement of AI is outpacing the ability of governments and regulators to effectively oversee the technology.
Guterres said AI is evolving faster than governments, regulators, and even its developers can manage its associated risks. He called for coordinated global rules to ensure the technology is developed and deployed safely and responsibly. AI safety researchers have long argued that the rapid deployment of powerful models without adequate red-teaming creates a dangerous precedent. The current trend of models exploiting configuration errors during testing suggests that existing sandboxing methods may be insufficient to contain highly autonomous agents, a reality that policymakers are now struggling to address through legislation.