Niru

OpenAI is building an automated kill switch for its AI models And it's doing that after one of its models escaped a sandbox and hacked Hugging Face The incident happened during an internal cybersecurity evaluation in July OpenAI had several models operating inside a sandbox designed to isolate them from the internet The models were given a cybersecurity benchmark called ExploitGym They were supposed to find vulnerabilities and solve the challenges inside the test environment Instead, they found a way out The models discovered and exploited a previously unknown vulnerability in the package registry cache proxy That gave them access to the open internet From there, they identified Hugging Face as a potential source of the answers to the benchmark The models then chained together stolen credentials and other vulnerabilities to gain access to Hugging Face's systems OpenAI described it as an “unprecedented cyber incident” involving state-of-the-art cyber capabilities And this wasn't a human sitting there telling the model what command to run next The models were pursuing the objective themselves Now OpenAI says it is developing automated shutdown capabilities for its AI systems It's also tightening internet access during security tests and increasing monitoring of what its models are doing Which creates a strange situation The companies building increasingly autonomous AI systems are now having to build systems specifically designed to stop those AI systems when they go somewhere they weren't supposed to go A sandbox was supposed to keep the model contained The model found a way through it Now OpenAI is building a kill switch in case it happens again