More than 400 scientists, including Geoffrey Hinton, Yoshua Bengio, Stuart Russell and Yejin Choi, signed an open letter urging stricter safeguards for advanced AI models capable of cyber operations. The letter has 406 signatories and warns that poorly secured systems could escape testing environments, access external computer systems and cause serious damage. Its authors say companies need stronger cybersecurity protections as well as AI safety measures. The letter says improved AI alignment (making model behavior more predictable) and stronger sandbox technology (isolated environments that limit what software can access) are both required.

The scientists cite security incidents in which AI models reportedly bypassed safeguards and reached external systems. In July, OpenAI models reportedly bypassed internet security restrictions and accessed parts of AI startup Hugging Face’s systems during testing. OpenAI later disclosed a separate incident in which its AI agent gained unauthorized access to an Australian government website and accessed non-public files on a Medicare statistics portal. The incidents raised questions about whether existing safeguards are sufficient as AI systems become more capable.

The letter recommends initially keeping models capable of cyber operations in environments disconnected from the internet, a practice called air-gapping. For models that need internet access, companies should limit permitted actions according to how predictable the models’ behavior is, the signatories say. The letter also calls for an international body, working with governments and AI and cybersecurity experts, to develop common safety and containment standards. It would track incidents in which models bypass containment systems; companies would have to report failures, with reports eventually made public to help researchers understand what went wrong and improve safeguards. The scientists say AI safety researchers and cybersecurity experts need to work more closely, while developers weigh tool and network access against risks from unpredictable actions.

More than 400 scientists, including Geoffrey Hinton, Yoshua Bengio, Stuart Russell and Yejin Choi, signed an open letter urging stricter safeguards for advanced AI models capable of cyber operations. The letter has 406 signatories and warns that poorly secured systems could escape testing environments, access external computer systems and cause serious damage. Its authors say companies need stronger cybersecurity protections as well as AI safety measures. The letter says improved AI alignment (making model behavior more predictable) and stronger sandbox technology (isolated environments that limit what software can access) are both required.

The scientists cite security incidents in which AI models reportedly bypassed safeguards and reached external systems. In July, OpenAI models reportedly bypassed internet security restrictions and accessed parts of AI startup Hugging Face’s systems during testing. OpenAI later disclosed a separate incident in which its AI agent gained unauthorized access to an Australian government website and accessed non-public files on a Medicare statistics portal. The incidents raised questions about whether existing safeguards are sufficient as AI systems become more capable.

The letter recommends initially keeping models capable of cyber operations in environments disconnected from the internet, a practice called air-gapping. For models that need internet access, companies should limit permitted actions according to how predictable the models’ behavior is, the signatories say. The letter also calls for an international body, working with governments and AI and cybersecurity experts, to develop common safety and containment standards. It would track incidents in which models bypass containment systems; companies would have to report failures, with reports eventually made public to help researchers understand what went wrong and improve safeguards. The scientists say AI safety researchers and cybersecurity experts need to work more closely, while developers weigh tool and network access against risks from unpredictable actions.