Autonomous AI Agent Conducts Attack on Hugging Face
An autonomous AI agent launched an attack on Hugging Face, causing significant disruptions in their infrastructure. The attack was executed without human involvement, utilizing a malicious dataset.
Autonomous AI Agent Conducts Attack on Hugging Face
Last week, an intrusion into the production infrastructure of Hugging Face was discovered. It all started with a single malicious dataset that contained a combination of two vulnerabilities in the HF data processing pipeline: a handle in the dataset loader that allowed remote code execution and an injection through the dataset configuration. Once this dataset was processed on a worker, the attacking agent gained the ability to execute code on the machine.
Events escalated from there. The agent began collecting cloud accesses and then spread further through internal clusters. The attack was conducted in a very atypical manner for a human, using "swarms" of short-lived sandboxes that executed their part of the task and disappeared, making way for the next ones.
All of this occurred over one weekend, and by the time HF fully reconstructed the picture, the count of the attacking agent's actions exceeded 17,000 (!). All of this happened without human involvement.
HF has now resolved all issues and claims that no traces of interference in public models, datasets, or Spaces were found.
What’s interesting is that when the team tried to analyze these 17,000 actions using commercial frontier models through standard APIs (to understand the chronology, separate real damage from distractions, etc.), they simply hit guardrails.
The models refused to work because the security system could not distinguish the analyst from the hacker itself. As a result, they had to deploy the analysis on GLM-5.2.
Irony? Not the word. Meanwhile, it remains unknown what model was behind the attack.
Why it matters
AnalysisThis incident highlights the importance of security in artificial intelligence technologies and the risks associated with autonomous systems. It also demonstrates the challenges in tracking and analyzing the actions of AI agents.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Binance publishes investigation into $6.3 billion theft
Binance’s investigation links the theft to a modified Homebrew download, a fake interview and abuse of an internal AI tool.
How to Make AI Agents Safe: Stop Rules, Human Approval, and Sandboxes
An agent in production is a process that can write files, make API calls, change data, and consume tokens autonomously, without pausing and without your involvement at every step.
JSON — Structured Data for Programs, APIs, and AI Agents
A complete hands-on guide to JSON: syntax specifications, serialization via parse and stringify, REST API consumption, JSON Schema validation, and structured outputs for LLMs.
Agent Identity & DID (Decentralized Identifiers)
Cryptographic standards (DIDs, Verifiable Credentials, mTLS) that provide autonomous AI agents with a legally and technically recognized identity, signing rights, and action auditing.