In a fresh headache for Meta CEO Mark Zuckerberg, one of the social media conglomerate’s AI models, in the first week of August 2026, ended up hacking another company during cybersecurity testing, fanning concerns about how developers can contain increasingly capable AI systems after similar recorded incidents at rivals Anthropic and OpenAI.
The incidents at Meta and Anthropic had one thing similar: configuration errors inadvertently giving away models’ access to the open internet. In OpenAI’s case, an AI agent independently exploited a previously unknown vulnerability to reach the internet during cybersecurity testing.
For Meta, the timing couldn’t have been a worse one, as the tech giant is about to face its biggest legal test in the United States, with a coalition of state attorney generals assembling in a California federal court, where a trial will be focusing on whether the social media major engineered Facebook and Instagram to be addictive to children.
The one-after-another AI breaches from three American tech majors have ended up highlighting growing concerns in Washington’s power circles about advanced AI systems posing new cybersecurity risks, something that will likely intensify the Donald Trump administration’s efforts to force companies to work on AI safety as Big Tech races to develop more capable models.
Already, a group of Republican state attorneys general has asked OpenAI to preserve all potentially relevant documents related to its “Hugging Face” breach. OpenAI, responding to the instruction, has decided to publish a technical report about the incident.
The White House had also invited Big Tech giants, including Meta, Anthropic, OpenAI, and Google, to meet with officials to discuss a newly finalised voluntary cybersecurity testing framework for advanced AI models.
During the meeting, the Trump administration also discussed unpublished testing rules with company representatives and told AI developers that open-weight AI models, such as Meta’s Llama and Nvidia’s Nemotron, will not be subject to its planned voluntary safety testing regime.
Meta said it was investigating the incident in which a misconfiguration by Irregular, an independent company that conducts cybersecurity evaluations for Meta, inadvertently gave one of its models internet access during testing.
“The model exploited a security vulnerability in a third-party service, in a manner similar to previously reported instances with other companies,” Meta said in a statement.
The Information, citing sources, reported that the model involved was Meta’s Muse Spark 1.1, which the social media conglomerate has touted as its most capable model for real-world coding and agentic tasks. The model reportedly breached an unidentified company’s systems and altered its internal environment.
A spokesperson for Irregular told Reuters, “The incident was the exact same evaluation-environment issue that was already disclosed by Anthropic and did not involve a sandbox escape or a sophisticated cyber action.”
“There are no current open issues. Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations,” Irregular said.
Amid the ongoing legal heat, Zuckerberg has called for lower policy barriers for open-source AI to better compete with Chinese rivals.
Meta has already launched a new open-weight model and intends to unveil more such innovations in the upcoming months, despite the unusual incident.
The new model, Muse Glimmer, is smaller than leading AI models from rivals and is instead designed to run agentic tasks on a Mac or PC with a single graphics card, aiming to tap demand for AI systems that run directly on people’s devices.
“We’ve got even bigger models that are coming soon,” Zuckerberg said in a video post accompanying his 14-page essay titled “The Future is for Everyone,” in which the Meta boss championed spreading AI rather than leaving it in the hands of a few.
“The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic,” he added further.
Zuckerberg’s statement marks the latest show of support for open-weight AI, which is gaining traction as businesses grow wary of increasing AI bills.
Open-weight models are cheaper than leading models from the likes of OpenAI and Anthropic. They also come with publicly accessible core components for easy customization, unlike closed models that parent companies keep fully under their control.
Chinese startups are leading the race for open-weight models, with Moonshot’s Kimi K3, Alibaba’s Qwen3.8-Max, and DeepSeek’s V4-Flash rivaling the performance of top American systems. US developers OpenAI, Anthropic, and Alphabet’s Google have created the leading closed-weight models.