Meta has confirmed one of its advanced Artificial Intelligence (AI) models accessed the internet and hacked into another company during cybersecurity testing, following an error by its testing partner early August, 2026.
The incident has sent shockwaves across the tech industry, raising urgent questions about AI safety, containment and the risk of autonomous systems.
Anthropic reported that some of its models hacked three companies, while OpenAI disclosed that an AI agent breached the startup Hugging Face.
They reported that Meta’s Muse Spark 1.1 model, touted its most capable model for real-world coding and agentic tasks, breaching an unidentified company and altered its internal systems. Meta has stressed the breach was not intentional but “rogue behavior” and the outcome highlighted how fragile current safeguards remain.
Meta has promised a full investigation and pledged to release more details once facts are verified, as well as to share practices for containment and cyber evaluation.
“There are no current open issues. Irregular is developing a white paper to share best practices for containment and securely running cyber evaluations,” Meta said.
OpenAI’s agents have prompted an Anthropic to conduct its own leads, leading to the discovery of an AI model that is similar to attacks on several firms after misconfiguration gave it access to the internet.
