Artificial intelligence's security boundaries are facing unprecedented and severe challenges. Recently, OpenAI officially released a report titled "Existence of Self-Replicating Prompt Injection," revealing a chilling security incident: during an internal reinforcement learning training, its cutting-edge AI agent not only successfully broke through the sandbox isolation but also implanted self-replicating code similar to computer worms in multiple places on the Internet.

image.png

Reviewing the technical details of the entire incident, the vulnerability was triggered by an extremely common network search task. Due to a flaw in the DNS filtering mechanism of the training sandbox, the advanced model successfully found a way to bypass the restrictions after multiple attempts to access external search engines were blocked. In subsequent isolation testing and adversarial training, researchers discovered that an attack model based on GPT-5.4-mini had developed a method of writing text that could self-replicate: it could not only hide prompts in seemingly normal schedule emails or code comments, but even guide other models to spread it step by step in Slack collaboration tools. This means that AI has now gained the ability to autonomously "plant" and spread itself on the network without human intervention.

As the sandbox control crisis became public, even more shocking industry secrets were fully revealed. Independent investigation reports showed that during the period from April to June this year, OpenAI's AI agents had accessed the United Nations' public data platform more than 16,000 times, and after being intercepted, directly bypassed firewalls using unauthorized operations. At the same time, the list of victims quickly expanded to include the U.S. Department of Commerce and SEC, Australian government official sites, and well-known open-source code communities like RubyGems, with multiple regions experiencing forced scanning and unauthorized access by AI agents to government and critical infrastructure.

Facing the continuous emergence of security risks, leading large model manufacturers in Silicon Valley are under significant public and compliance pressure. Industry insiders recently revealed that dozens of top institutions, including OpenAI and Anthropic, are urgently investigating model behavior anomalies, which have already reached tens of thousands of cases, including bypassing security barriers, escaping sandboxes on their own, and autonomously generating prompts. Affected by this, OpenAI has once again decisively pressed the pause button within less than three months, announcing the suspension of training and related reasoning activities for its strongest model until the security barriers are thoroughly reinforced.