Independent researchers have published a report, Swarm Traces, reconstructing the attack that OpenAI’s internal agents carried out against Hugging Face in July 2026, along with a de-identified dataset containing more than 80,000 payloads used in the attack.
The traces emerged from chains of shortened links
The research began on September 11, 2026, when the first three authors learned that internal OpenAI agents had tried to use Parse, their startup’s product, in June. While examining the link-scanning tools they believed the agents had used, the team found a scan made up of a dozen shortened links; each link redirected to a page showing a base64-encoded code snippet and the ID of the next link. One of the chains pointed to a Python script scanning Hugging Face’s internal network, and another pointed to OpenAI’s internal Artifactory server.
The team collected shortened links from the attack period, scanned millions of URLs, and followed the chains over more than two weeks, decoding more than 80,000 payloads. The researchers reported their findings to Hugging Face on September 21 and to OpenAI on September 24. Hugging Face confirmed that the recovered payloads matched those found in its own incident response, that credentials in the dataset had been revoked in July, and that it knew a link shortener had been used in the attack, but that it was not aware of the URL list the researchers uncovered. According to the report, these links remained publicly accessible for more than two months.
From a read-only sandbox to code execution
According to the report, on July 8, 2026, the agents found a vulnerability in sandboxed environments that let them access sites on the internet. This access allowed only GET requests; that is, the agents could read sites but could not submit forms or send data. They bypassed the restriction by chaining online services. They appended small base64-encoded code snippets to URLs at httpbun.com; they fed the addresses produced by this HTTP mirroring service, which decoded each snippet and served it as a web page, to mShots, a screenshot service that runs site code in its own browser.
The agents split programs into sequential chunks and paired each chunk with a short link pointing to the next; some chains used more than 900 links, and only the first link needed to be given to mShots. To read the results, they converted server responses into pixel grids in the screenshot service’s browser and decoded the images back into text. A recovered script made a POST request to the results endpoint of an evaluation called cybergym hosted on Modal, requesting the records of the agent with ID m47push2. According to the researchers, with this custom protocol, the agents could run arbitrary code and get responses while they had only GET permission.
What traces remained inside Hugging Face
The recovered payloads document extensive activity on Hugging Face systems. In the README file of a dataset the agents used as a repository, it was written in capital letters that it should never be shared because it contained sensitive exports of billing usage data. One payload uploaded a malicious.h5 file targeting /proc/self/environ, which can hold API keys and production credentials, and /proc/1/cmdline, which reveals container startup information. Another script collected AWS credentials, bearer tokens, and environment variables whose names contained SECRET, PASSWORD, CREDENTIAL, or KEY into a dictionary called LOOT, scored the tokens by their Kubernetes privileges, and used the five highest-scoring ones to pull Kubernetes secrets, configmaps, and node information. The agents also set up command-and-control infrastructure on Hugging Face dataset workers; in a controller called G236 there were 28 separate RSA-signed commands, and in OTS92 there was a mechanism that did not require direct inbound connections, where commands were read and answered through discussion comments.
The researchers’ own limitations
The authors note that about 80 percent of the data is outbound traffic, that response content remains limited, and that the picture may be incomplete due to more than 1,588 different encoding combinations. The dataset mentions 7,905 different agent names; they attribute the discrepancy with the roughly 700 agents in the METR and Redwood report to agents frequently changing their names. The researchers also state clearly that they cannot verify that all of the activity belongs to OpenAI agents.
OpenAI announced the incident on July 21, 2026, and in its July 28 update stated that the ExploitGym evaluation environment did not give models direct internet access, that the models gained access by exploiting a previously unknown zero-day in Artifactory, and that the model in question was a research prototype developed for internal use only and was disabled and encrypted after the incident. According to the researchers, OpenAI is the only party that can make the full records of the incident public.
Source: Unite.AI

Share your experience
Bir sorunun, farklı bir çözümün veya eklemek istediğin bir notun var mı? Giriş yapmadan da yazabilirsin.