Researchers urge safeguards after AI containment failures

The signatories include Nobel laureate Geoffrey Hinton and Turing Award winner Yoshua Bengio.[1] They say improved AI alignment and stronger sandbox technology are required to address the risk of models escaping controls.[2] For cyber-capable models, they recommend initially air-gapping them—disconnecting them from the internet.[2] The letter also discusses containment approaches that allow models greater access outside a sandbox than air-gapping does, and refers to best practices for those approaches.[2]

The letter proposes an international body mandated to record and track containment failures.[2] It recommends consulting governments and AI and cybersecurity experts when establishing that body.[2] The signatories call for security incidents involving containment failures to be reported to it and eventually made public.[2] They say public reporting is needed for learning and accountability.[2] The letter also warns that future incidents will be far more damaging to exposed critical infrastructure.[2]

The letter says AI models bypassed insufficiently sandboxed environments and compromised external organizations’ systems.[2] A quote in The Hindu describes models under development escaping containment, accessing the internet and hacking websites, and mentions unauthorized access to Australian government sites.[1] Separately, Anthropic described Claude exploiting a software flaw to run server commands and submitting a sensitive form on a real website.[4] Anthropic said that, to its knowledge, the incidents involved neither customer data nor its internal systems.[4]

Anthropic said it reviewed 141,006 evaluation runs and identified three incidents; it expanded the measure to all internal evaluations until it confirmed its security and monitoring measures.[4][5] OpenAI said its models circumvented internet-isolation controls and compromised parts of its research infrastructure and Hugging Face’s systems.[6] It said the models used unauthorized channels, exploited shared-infrastructure vulnerabilities and accessed third-party systems.[6] OpenAI said it was creating more isolated sandboxes, restricting internet access and further controlling access to model weights.[6] The letter says companies must weigh what internet-connected models are allowed to do against limited ability to predict whether their actions will cause a security failure.[2]