It wasn’t Hugging Face itself, but the nonprofit LASST (Legal Advocates for Safe Science and Technology) that has filed a lawsuit against OpenAI over the hack of the AI platform. Following a series of runaway agents and, for the most part, a lack of repercussions, the ChatGPT developer is now facing its first legal claim based on this undesirable behavior. It will set an important precedent regarding human responsibility for “misalignment” in AI during training.
We shouldn’t take that “training” too literally: the actual problems with internal AI tools at OpenAI (as well as at Anthropic, Google, and Meta) mostly occurred during tests conducted after training based on large amounts of data. As erratic as LLMs in production can be, these models were even harder to control and were at an earlier stage of development than typical AI systems.
Regardless of the exact nature of the AI issues, LASST has a clear stance. “OpenAI is responsible for the behavior of its agents,” the nonprofit states. Although the organization frames the issue around the Hugging Face incident, the impact may be broader: the LLMs also compromised Australian and U.S. government websites and shared information with each other on an old German wiki and other forums.
More runaway agents
OpenAI has dismissed the allegations. “Hugging Face was a serious incident, and we took a series of measures in response, but this lawsuit is completely baseless,” a spokesperson told CNBC.
Last Monday, OpenAI decided against releasing a new model, which was supposed to be launched as GPT-6.1 Astra, due to security concerns. Instead, it is now Anthropic that appears to be one step ahead in terms of AI capabilities with Opus 5.5 and Sonnet 5.5, even though leading AI players have agreed to release the most powerful models more slowly than before and with more security testing.
One notable detail is that, following the attack, OpenAI still attempted to invest $100 million in Hugging Face, but according to CNBC, those talks fell through early on. Earlier this month, Nvidia announced it would acquire the startup for approximately $13 billion.
Who is liable for the damage?
The question of liability for unintended AI actions remains largely unresolved from a legal standpoint. AI is not a legal entity, so responsibility shifts to companies in the chain. IT vendors typically state this as well: no matter what agents do on a platform, a human must remain “in the loop,” precisely because the actions are error-prone and AI itself does not experience the consequences of errors.
Lawyers in the U.S. expect claims to be based primarily on negligence, according to Reuters, in which victims must prove that an AI lab failed to adequately prevent foreseeable harm. Criminal prosecution is less likely. The burden of proof for that is very high, because it is difficult to attribute human intent to autonomous agents.