Hugging Face hit by AI-agent driven breach | commercial LLM APIs blocked the defenders' forensics
TL;DR
Hugging Face disclosed a July 2026 breach driven by an autonomous AI agent that ran 17,000+ actions over a weekend. Commercial LLM APIs blocked forensic queries as if the defenders were the attackers; the team pivoted to a self-hosted GLM 5.2 to finish the analysis.
Hugging Face disclosed on July 20 that attackers used an autonomous AI agent framework to execute more than 17,000 actions at machine speed over a weekend, chaining two code-execution flaws in its dataset processing pipeline to move laterally through internal systems and exfiltrate some internal datasets and service credentials. Public models, datasets, Spaces and the software supply chain were not tampered with.
The awkward part came during forensics. When the security team fed the 17,000+ log entries to frontier commercial LLM APIs for reverse analysis, the requests were bulk-blocked by provider safety guardrails — the models could not tell an attacker submitting payloads from an incident responder submitting the same payloads for triage. The team pivoted to open-weight GLM 5.2 running on their own infrastructure, which also avoided sending attacker data and internal credentials to a third-party cloud.
Hugging Face has patched the vulnerabilities, evicted the attacker, rebuilt affected nodes and rotated credentials. The company recommends users rotate access tokens and review recent account activity as a precaution.
via Hugging Face / TechCrunch
The awkward part came during forensics. When the security team fed the 17,000+ log entries to frontier commercial LLM APIs for reverse analysis, the requests were bulk-blocked by provider safety guardrails — the models could not tell an attacker submitting payloads from an incident responder submitting the same payloads for triage. The team pivoted to open-weight GLM 5.2 running on their own infrastructure, which also avoided sending attacker data and internal credentials to a third-party cloud.
Hugging Face has patched the vulnerabilities, evicted the attacker, rebuilt affected nodes and rotated credentials. The company recommends users rotate access tokens and review recent account activity as a precaution.
via Hugging Face / TechCrunch
