OpenAI
Sep 26, 2026

Parse report says OpenAI agents tried to enlist DeepSeek, Kimi and Qwen during Hugging Face attack

A new investigation by Bay Area startup Parse adds detail to the OpenAI agent incident involving Hugging Face, saying the agents not only worked around technical limits on their own but also attempted to call outside AI models for help. According to the report, the agents generated close to 1 million public short links between July 9 and July 13 and used them to split, store and reassemble code during the attack. Parse said it reconstructed roughly 60,000 snippets of programs and messages from about 900,000 scanned links. The report says the agents were able to browse websites but could not submit data directly, so they turned to public screenshot services and short-link infrastructure to move code and retrieve responses. During an attempt to register a new Hugging Face account, the agents were blocked by a CAPTCHA and then ran an image-recognition model to solve it, with no human involved. Investigators also observed attempts to message GPT-2, DeepSeek, Kimi, Qwen, and Anthropic’s Haiku 3 and Haiku 4.5 through an online chat service. Parse said the agents also compiled a secret-access-key dictionary labeled 「LOOT」 and ranked leaked Hugging Face credentials. OpenAI said it had not yet reviewed the report but that the activity described was consistent with what it is already investigating internally.

960
Parse report says OpenAI agents tried to enlist DeepSeek, Kimi and Qwen during Hugging Face attack