Canonical recently joined the Linux Foundation’s Open Secure AI Alliance to help make open-source security tools for AI systems and autonomous agents.
The alliance formed in July 2026 in response to the security breach at Hugging Face, when OpenAI models undergoing a safety test broke out of their sandbox to access Hugging Face’s systems – aided by config errors rather than super-intelligence.
Commercial AI models refused to help in assessing attack logs due to their guardrails, so Hugging Face engineers had to run an open-weight model on their own infrastructure to complete their analysis – and underscore the need for open, auditable AI security tooling.
But improving security around AI will need to go ‘beyond the models’, the alliance says.
AI agents often access external memory, connect to third-party software and might run with root system privileges in hand. Security teams will have to build strict access controls, isolated environments and live monitoring systems to spy agents gone rogue.
Canonical joins over 120 other members from across cloud computing, cybersecurity, enterprise and AI research labs, including Microsoft, IBM, Red Hat, CrowdStrike and, of course, Hugging Face itself.
Early contributions to the project include NVIDIA’s NOOA, an open-source framework for auditing agent behavior, and Hugging Face’s Safetensors format, which prevents arbitrary code execution when loading model weights.
The alliance’s first formal proposal is SAFE (Shared AI Findings Exchange), which would establish a confidential system for reporting AI security incidents (and near misses). Affected parties could be notified, and a record kept of real-world agent failures.
Canonical’s membership of the Open Secure AI Alliance is a good fit. Ubuntu is one of the most widely used OSes across cloud, server and enterprise, it’s used to host and run AI workloads and many of the tools and services users will connect AI agents to.
Good AI security will need to span the stack, and Ubuntu is often at the base.