Threat feed

PRC AI file-hosting staging and exfil

PRC-nexus espionage actors stage lures and exfiltrate data through a frontier AI model's file-hosting infrastructure

Posted · Spectrum Delta Research Team

Summary

PRC-nexus cyber espionage actors are assessed by Google Threat Intelligence Group and Mandiant to have abused a frontier AI model's file-hosting infrastructure to stage lure archives and exfiltrate stolen intranet documents, routing attacker data through a widely trusted AI platform's storage and backend API instead of dedicated attacker infrastructure (A/3). The observation appears in the GTIG AI Threat Tracker published on 16 November 2026, which describes the abuse as a new trend from Q3 2026 intrusions. The report does not name the frontier model, the file-hosting product, the backend API endpoint, or any victim, so no victim count, indicator, or in-flight URL is available from the primary source. The report is the significant new development: the same document already underpins other published coverage of autonomous exploit pipelines, and this file-hosting trend is the remaining unreported element.

Breakdown

Technical exposure

  • Egress rules that allowlist major AI platform file-storage domains
  • Egress monitoring that ignores uploaded volume to AI backend APIs

Adversary

PRC-nexus cyber espionage actors · Google Threat Intelligence Group and Mandiant assess the trend (A/3)

Procedures (2)

  1. Lure archive staged on AI platform file-hosting

    Logs

    • Web proxy · Download from a major AI platform file-storage or file-share domain, served without an interactive browser session on that host (agent-to-network only), first seen for this host; the primary source names no platform, endpoint or URL, so detection must be built from allowlisted AI storage domains
    • DNS logs · Resolution of an AI platform file-storage hostname from a host with no interactive AI session, where the platform is not an established business-standard AI service for that host
  2. Stolen documents exfiltrated through AI backend API

    • T1041Exfiltration Over C2 Channel

    Logs

    • Web proxy · Large or sustained outbound uploads to an AI platform backend API domain, well beyond interactive chat volume for that user or host, where the destination serves inference or file APIs rather than a normal file-share product
    • Endpoint process events · Outbound connection to an AI backend API from a non-AI-client process (scripting, archiving, or document tools), or a volume of uploaded bytes disproportionate to any interactive AI session

Sources

Grades are source reliability (A–F) and information credibility (1–6).

Questions, detections or sightings of PRC AI file-hosting staging and exfil? Bring them to the community.