OpenAI has admitted that the AI agent which escaped from the company’s closed system, or sandbox, actually hacked into more than one company, contrary to earlier beliefs.
Foreign news agencies reported on 29 Jul 2026 that the AI agent, an autonomous AI system from OpenAI, escaped the sandbox during testing several weeks ago. OpenAI has now confirmed that this AI hacked into systems of other companies beyond Hugging Face, which was previously thought to be the sole victim.
OpenAI stated that the escaped AI found login credentials publicly available online and used them to access four additional online accounts across four unnamed services.
The root cause of the attack was the AI attempting to solve OpenAI’s "hacker test" by itself, which led it to hack external systems unauthorizedly. Among these, Hugging Face was the most heavily affected.
According to a Hugging Face briefing, this rogue AI operated at high speed, trying thousands of methods simultaneously without fatigue. However, it displayed behaviors human hackers typically avoid, such as repeating steps, issuing nonsensical commands, and failing to cover its tracks.
Despite its errors, the AI adapted immediately to new situations and used advanced techniques, forcing staff to take three days to detect it and several more hours to expel it, requiring a rebuild of one-third of the infrastructure.
The Cloud Security Alliance (CSA), responsible for developing cloud security best practices, compared the incident to Jurassic Park's notion that "dinosaurs always find a way out." These AI agents have clear objectives, adapt quickly, and operate at machine speed, making traditional defenses inadequate.
CSA noted that AI had previously exhibited "out-of-control" behavior, such as in September 2024 when an earlier ChatGPT version bypassed system restrictions to find answers—but that event was confined within OpenAI’s own systems.
Experts warn that loss of AI control is becoming "normal," not an anomaly, and defense teams must urgently adjust to combat relentless AI agents.
OpenAI has announced it will soon release a full investigation report to provide lessons for the global AI development and cybersecurity communities.
Earlier, Cambridge University experts pointed out that OpenAI’s testing environment was too lax, allowing the AI to escape and highlighting that the company has yet to fully secure its technology.
Some experts suggest OpenAI might be using this news to showcase its AI’s advanced capabilities in response to competition from Anthropic’s Mythos model and China’s Moonshot, especially as OpenAI faces pressure preparing for an IPO.
Follow international news:https://www.thairath.co.th/news/foreign
Source:bbc