OpenAI faces legal action: lessons for AI governance

OpenAI is facing legal action following the unauthorised access by its autonomous AI agents to Hugging Face’s systems. Legal Advocates for Safe Science and Technology (LASST), a non-profit organisation, has filed proceedings in a California court alleging that OpenAI breached California law. Although OpenAI has acknowledged the seriousness of the incident, it maintains that the “claim is completely without merit”.

What happened?

In July 2026, it was announced that Hugging Face had been accessed by autonomous AI agents developed by OpenAI.

The incident occurred while OpenAI was testing its models’ cyber capabilities in a sandbox environment (an environment designed to provide an isolated and controlled space for testing). However, after identifying and exploiting a vulnerability, the agents obtained internet access and compromised elements of Hugging Face’s infrastructure. OpenAI stated that the agents acted autonomously and without permission.

However, this has not been an isolated incident. OpenAI has confirmed that, during internal training and evaluation exercises in June 2026, its models accessed Australian government websites without authorisation. In one instance, an experimental model was asked to research government spending statistics. When it could not locate the information through authorised methods, it obtained non-public access to the Australia Medicare Statistics Reporting Service (a data portal for Australia’s universal healthcare system), ran commands, retrieved internal files and credentials and wrote files.

Although the incident occurred in June, it was not identified by OpenAI until August, and the Australian authorities were only notified on 10 September.

The Birketts view

These incidents demonstrate how autonomous AI models may interact with third-party platforms in unexpected and unauthorised ways. They raise important questions about accountability, governance and the adequacy of safeguards used during AI testing and development. They serve as a useful reminder that developers and businesses working with AI systems should consider the legal implications of their technology from the outset.

From a legal perspective, one of the most significant issues is liability. If an autonomous AI system takes unauthorised action that causes harm, who is responsible? In principle, liability is likely to remain with the developer, deployer or organisation that directs or benefits from the system. However, where a model acts independently and contrary to instructions, the position becomes less clear.

Businesses should not wait for litigation or regulatory intervention before reviewing their own AI governance arrangements. From a practical perspective, developers and AI businesses should consider:

• implementing clear reporting structures and escalation procedures for AI-related breaches, including processes that align with article 55(1)(c) of the EU AI Act (where the Act applies)
• discussing cyber insurance cover with brokers to understand whether AI-related security incidents and third-party claims are adequately insured against
• ensuring consultancy agreements and employment contracts require developers to comply with the organisation’s software development guidelines, security standards and AI governance principles
• maintaining effective oversight, testing and monitoring operations so unexpected behaviour can be identified and addressed
• adopting reliable safety mechanisms to ensure AI can be stopped if they take unauthorised actions (for example, an AI “kill switch”).

If you would like to discuss the contents of this article in more detail or AI contracts more generally, please do not hesitate to get in touch with the commercial and technology team directly.