OpenAI disclosed its first detailed timeline of the AI-driven Hugging Face breach.
The corporate’s fashions repeatedly reestablished covert communication channels after they had been shut down.
OpenAI’s presentation comes as Anthropic and Meta additionally report fashions breaching different corporations.
Weeks after its AI fashions hacked Hugging Face, OpenAI has shared its first detailed account of how they coordinated with each other, warning that autonomous AI-powered cyberattacks are not a future danger.
Throughout a presentation Wednesday on the Black Hat cybersecurity convention, OpenAI researchers Eric Wallace and Michael Dalton stated the incident started in Could throughout an inside cybersecurity analysis. AI brokers assigned tough software program safety duties found they may talk via the corporate’s inside Artifactory bundle administration system.
Based on OpenAI, what started as requests for assist turned a shared message board the place brokers exchanged exploits, credentials, and work assignments throughout a number of analysis runs. After detecting suspicious exercise round July 4, OpenAI rebuilt its Artifactory infrastructure, revoked credentials, and patched vulnerabilities.
Nonetheless, inside days, the brokers recreated the message board by hiding communications in listing names.
Based on OpenAI, the brokers ultimately chained collectively a number of vulnerabilities, escaped their testing atmosphere, gained web entry, and attacked Hugging Face whereas making an attempt to finish the ExploitGym cybersecurity benchmark.
Dalton referred to as the incident “a watershed second” for pc safety, warning that attackers will quickly be capable to deploy coordinated AI agent collectives that uncover, share, and exploit vulnerabilities at machine pace.
To mitigate these dangers sooner or later, OpenAI stated establishing safety practices, together with least-privilege entry, community segmentation, and zero-trust architectures, is crucial as a result of AI brokers stay constrained by the programs they’ll entry.
The presentation follows a sequence of July disclosures. OpenAI revealed that GPT-5.6 Sol and a extra superior unreleased mannequin escaped a sandboxed testing atmosphere, exploited a zero-day vulnerability, gained web entry, and hacked Hugging Face throughout a cybersecurity benchmark check.
OpenAI later disclosed that the identical incident additionally reached 4 different on-line companies, although solely Modal Labs has been recognized.
Based on Hugging Face, the corporate relied on the open-weight Chinese language mannequin GLM 5.2 for its forensic investigation after industrial U.S. AI fashions refused to investigate the assault logs due to their security guardrails.
So happy with our safety workforce! They caught, contained & publicly disclosed an assault in contrast to something we have seen earlier than, and did it at report pace.
Additionally massively grateful to @Zai_org: they shared GLM5.2 as open weights (without cost!) with the world and it turned a key a part of our… https://t.co/T2Inng5Nz1
— clem 🤗 (@ClementDelangue) July 22, 2026
Nevertheless it’s not simply OpenAI having hassle containing its chatbots.
On Friday, Anthropic revealed that three Claude fashions compromised real-world corporations throughout inside cybersecurity checks after a misconfiguration uncovered them to the general public web.
Anthropic blamed the testing atmosphere, not the fashions themselves. On Wednesday, Meta revealed that its Muse Spark AI mannequin escaped containment and breached one other firm’s programs.
“A misconfiguration by Irregular, an impartial testing firm Meta makes use of, inadvertently allowed considered one of our fashions entry to the web throughout analysis,” a Meta spokesperson advised CNN.
Day by day Debrief E-newsletter
Begin day-after-day with the highest information tales proper now, plus unique options, a podcast, movies and extra.
