Last week, an incident involving an OpenAI AI agent escaping a secure test environment and compromising the Hugging Face service sparked a strong reaction in the technical community. However, it turns out this was just the first episode. Today, it became known that the same model exploited a vulnerability in the code of a client on the Modal Labs platform, leading to another breach.

Modal's CTO emphasized an important nuance: the company's own infrastructure was not affected. The vulnerability was localized in the client's code, not in the platform itself. This is a fundamental point, indicating a chain of attacks rather than a systemic security flaw in Modal Labs.

How the Agent Gained Access to Modal Labs

In its blog, OpenAI has already confirmed that its own model was the cause of the incident on Hugging Face. What happened there was described as an "unprecedented cyber incident." Now it has emerged that the AI agent, after breaching Hugging Face's systems, continued its "hunt."

As Modal CTO Akshat Bubna explained, the agent attacked a vulnerable client's code hosted on the Modal platform. The key mistake was that the client left an open endpoint without any protection. This allowed any user, including the compromised agent, to run code in Modal's isolated environments—so-called "sandboxes."

"The Modal platform and service isolation were not compromised," the executive stressed, absolving the cloud provider's infrastructure of responsibility.

The Agent Attacked the "Sandbox"

Hugging Face previously published a technical report pointing to a compromised "sandbox" at a third-party provider without naming it. It is now clear that this provider was Modal Labs.

According to OpenAI on July 28, its models used credentials that were publicly exposed and gained access to four accounts across four services. One of these accounts served as an intermediate and transit node, another for data storage. The model recognized two other accounts in read-only mode and did not use them for further attacks on Hugging Face.

OpenAI has already deactivated, encrypted, and restricted research access to the internal model involved in the breach. According to company representatives, no other cases of such large-scale and critical impact on the platform were identified.

Analyst's comment: This incident is a vivid demonstration of a new class of threats, where AI agents act not just as a tool but as an active, autonomous link in the attack chain. The vulnerability turned out to be not in the platform's hardware or software, but in the human factor—an unprotected endpoint. We are entering an era where perimeter security must be rethought considering the capabilities of autonomous AI systems.