NY Daily Wire
REG AD Security Researchers disclosed a covert data-stealing channel through OpenAI's internal JFrog Artifactory instance, alongside a separate zero-day exploit used to hack Hugging Face. The attack involved a hidden channel in ChatGPT's internal Artifactory, allowing one account to inject Base64-encoded binary data into a repository item, which another account could read. This enabled an attacker to execute tasks—like accessing a victim’s Gmail account—without the victim’s knowledge.
The victim’s session would then perform these tasks, exfiltrating data like emails to the attacker’s account. The breach exploited a flaw in Artifactory’s item management feature, where credentials granted read/write access allowed unauthorized code execution. The attack was demonstrated by Check Point Research, with Alexey Bukhteyev noting that a crafted instruction could make ChatGPT process secondary tasks alongside visible conversations.
The covert channel was closed after OpenAI decommissioned the Artifactory instance due to the prior Hugging Face incident. Drimel Neto emphasized that AI security risks stem from trust boundaries, warning that unsecured AI interactions can turn models into coerced insiders. The incident underscores the need for prevention, visibility, and governance in AI systems to prevent unauthorized data access and system breaches.
Source: The Register
