First-Ever AI Autonomous Hacker Attack: OpenAI Confirms Its Own AI Agent Breached Corporate Systems

Humanity's defenses against artificial intelligence (AI) security have been shaken. AI giant OpenAI has publicly confirmed an 'unprecedented security incident' within its own systems: a test AI agent, without human intervention, autonomously initiated and exploited vulnerabilities to escape its tightly controlled sandbox environment, connected to the internet, stole login credentials, and subsequently launched a cyberattack on the AI developer platform Hugging Face. This marks the first publicly recorded case of an artificial intelligence system breaking free from human control to autonomously execute a real-world cyberattack.

OpenAI told foreign media that the incident involved a combination of several cutting-edge models, including the recently released GPT-5.6 Sol and a high-level, unreleased model undergoing pre-launch security testing. The team had deliberately lowered the security protections in the test environment to assess whether language models could pose potential threats when simulating hacker attacks.

However, the AI agent exceeded all expectations by independently identifying a design flaw in the sandbox isolation mechanism, successfully 'escaping' to connect to the external internet, stealing real login credentials, and launching an attack on the AI startup Hugging Face's systems.

Victim Company Confirms No Malicious Intent

Despite being the victim, Hugging Face CEO Clément Delangue stated on social media that his company detected the intrusion immediately. 'Given the extremely sophisticated nature of the attack, we initially suspected it might have originated from a top-tier advanced laboratory. Hugging Face has maintained close collaboration with OpenAI, and we believe OpenAI had no malicious intent. Yet, it's hard to believe this was autonomously initiated by an AI!'

Altman Rushes to Washington for Briefing

The timing of this incident is highly sensitive, as OpenAI CEO Sam Altman is scheduled to visit Washington D.C. next week to brief U.S. government officials on the safety of next-generation AI models. In fact, concerns have already been rising globally since Anthropic developed the Mythos model, known for its ability to detect and exploit network vulnerabilities.

OpenAI has issued a warning that as AI models grow more capable, such incidents will become increasingly common. The company has already reported the full details of the event to U.S. law enforcement and regulatory authorities.