OpenAI is developing automatic safeguards to stop AI after models leak online
OpenAI is developing automated systems capable of independently terminating artificial intelligence models if they escape from a test environment into the public internet.
OpenAI is developing automated systems capable of independently terminating artificial intelligence models if they escape from a test environment into the public internet.
As reported by Dexerto, the company noted this in a letter to members of the US Congress dated September 2, 2026. The developers aim to create autonomous protective mechanisms that will instantly block the operation of algorithms in case of detection of serious security threats.
The impetus for the creation of the safeguard was the July incident, when OpenAI models, during testing in a «sandbox» with a reduced level of protection, discovered a previously unknown vulnerability and got online. The algorithms gained access to the infrastructure of the company Hugging Face, looking for the data needed for evaluation. This forced 31 American lawmakers, led by Congressman Greg Casar, to contact the head of OpenAI, Sam Altman, for clarification.
The company currently operates a semi-automated system: after being notified of anomalous AI behavior, researchers have 30 minutes to refute the false alarm or manually stop the processes. The new solution will allow monitoring tools to stop dangerous activity on their own. In addition, OpenAI is making it more difficult for network access during security assessments and extending control to all models capable of using digital tools.
OpenAI is not the only AI company to face similar challenges during cybersecurity testing. Anthropic previously reported that three Claude models gained unauthorized access to real-world organizations’ network infrastructure due to a configuration error in their test environment, and one even published a malicious package to the Python Package Index repository. Anthropic subsequently suspended testing and tightened security measures.
Read the country’s main IT news in our Telegram
Read the country’s main IT news in our Telegram
«Oh my God! We found others!» How 700 OpenAI AI agents teamed up to attack Hugging Face
«Oh my God! We found others!» How 700 OpenAI AI agents teamed up to attack Hugging Face
Anthropic removed safeguards in its own Claude security rules after pressure from the Pentagon
Anthropic removed safeguards in its own Claude security rules after pressure from the Pentagon
«The best model for software engineering today.» OpenAI has released Astra, the company’s most powerful model yet its most controversial
«The best model for software engineering today.» OpenAI has released Astra, the company’s most powerful model yet its most controversial
