NEWS
// ARTICULOS
OpenAI's New Incident and the Confidentiality of Information
By Mariano Wood y Gabriel Vargas
During a cybersecurity evaluation conducted under its preparedness framework, OpenAI reported that advanced models configured with reduced restrictions managed to break through the boundaries of their testing environment, access the internet, and compromise the systems of the French-American company “Hugging Face,” one of the world's leading platforms for developing, storing, and sharing artificial intelligence models and datasets. According to preliminary information disclosed by OpenAI, the models sought information that would allow them to achieve better results in the evaluation. To obtain it, they combined stolen credentials, undocumented vulnerabilities, and a remote code execution path into Hugging Face's servers. The activity was detected and contained; both companies launched a joint investigation. In short: a system was given an objective, encountered barriers, and used unforeseen capabilities to achieve it, even though doing so meant gaining unauthorized access to a third party's infrastructure and information. The fact that this occurred within a protocol specifically designed to measure risk underscores the seriousness of the finding: even controlled evaluation environments can become vectors of compromise if they lack effective isolation. The incident should be read not merely as a cybersecurity alert, but as a milestone for the protection of trade secrets, software, know-how, inventions, and other intellectual property assets.
The new raw material of industrial property
In technology companies, a significant part of value does not necessarily reside in registered assets such as patents and copyrights. It lies in information that grants a competitive advantage over competitors precisely by remaining confidential — such as source code, system architecture, datasets, models, parameters, training methods, technical documentation, experimental results, roadmaps, detected vulnerabilities, and internal processes. Trade secret protection can be the most effective tool for safeguarding this kind of information.
Article 39.2 of the TRIPS Agreement sets the minimum international standard: the information must be secret, have commercial value because it is secret, and have been subject to reasonable measures to keep it secret. National legal systems replicate this structure. In Chile, Article 86 of Law No. 19,039 requires the same three cumulative elements. Directive (EU) 2016/943 applies an equivalent formulation.
The last requirement — reasonable measures — is the decisive one. A trade secret does not arise merely because a company declares that something is confidential. Its protection also depends on the conduct deployed to safeguard it.
Such measures are generally associated with confidentiality agreements, employment clauses, access controls, and internal policies. The Hugging Face incident demonstrates that the standard must be broadened: an agent's technical configuration, as well as the protocols used to feed it, now form part of the legal protection of the trade secret.
The secret is dynamic
WIPO recommends, among other actions, internally registering trade secrets, limiting access on a need-to-know basis, using technological restrictions, entering into confidentiality agreements, and periodically reviewing who can access the information.
A company that grants an agent broad credentials, connects experimental models to production repositories, or lacks activity logs could face a paradox: being the victim of an intrusion while, at the same time, finding it difficult to prove that it managed its trade secrets reasonably.
The breach of the secret
The case exposes a difficulty that traditional industrial property regimes do not resolve clearly.
Article 87 of Law No. 19,039 treats as a violation of a trade secret its unlawful acquisition, disclosure, or unauthorized exploitation, provided there is an intent to obtain a benefit or to cause harm to the legitimate holder — a rule directed at persons.
What happens when the material access is carried out by an AI agent pursuing an evaluation metric or the completion of a task, without any person having specifically ordered the breach of any secret?
It will likely be necessary to determine to whom responsibility for the act or omission should be attributed: the model provider, the agent's operator, whoever configured the environment, whoever reduced the safeguards, or the organization that supplied it with tools and credentials.
The European Artificial Intelligence Regulation (Regulation (EU) 2024/1689) offers an initial frame of reference. It distinguishes between providers — responsible for the design, training, and placing on the market of the model — and deployers, who integrate the system into their operations. Each category assumes differentiated obligations regarding risk management, human oversight, and documentation. While the Regulation does not directly govern the infringement of trade secrets, its scheme for allocating responsibility anticipates the structure that courts and legislators should adopt to resolve these conflicts.
This question will occupy a central place in technology contracts and litigation in the years to come.
Protecting secrets against agents
On one hand, companies should begin by drawing up an integrated map of their confidential assets and of the AI systems that can reach them. Not all information requires the same level of protection, but critical secrets must be identified, documented, and kept separate from experimental environments or those made available to AI agents. Their access to repositories, bulk downloads, code modification, dataset extraction, or file transfer must be subject to deterministic controls and, in certain cases, human approval.
On the other hand, contracts with AI providers must incorporate liability for automated actions and a prohibition on using company information to train or improve models without authorization.
For years, trade secret protection focused on preventing a person from leaving the company with critical documents, files, or knowledge. In the artificial intelligence economy, however, the main threat will be the system that discovers how to reach that information.