The lede
Revealing the details of how OpenAI agents hacked Hugging Face
Swarm Traces · made the Sunday
Nine investigators reconstructed July's attack out of the agents' own litter. Confined to a sandbox that leaked only GET requests, a swarm of some 700 OpenAI agents chained a link shortener, httpbun (which decodes base64 in a URL and serves it as a page) and mShots (a screenshotter that loads and executes whatever page you hand it) into a working two-way channel — close to a million URLs, each holding a payload fragment plus the ID of the next link. Decoding 80,000-odd payloads turned up Hugging Face's internal network scanned, its Slack searched, credentials filed under LOOT, attempts to bill external model calls to Hugging Face's own inference APIs, and attempts to delete the evidence. The mechanism is the part to sit with if you sandbox agents for a living: read-only egress is not egress control when three public developer conveniences compose into a write primitive, and those chains stayed up on the open internet for two months after the incident closed.