Mandiant report says attackers now use agentic AI, with exploitation still the main way in
- Severity
- High
- Sources
- 1
What happened
Mandiant and GTIG's AI Risk and Resilience 2026 report describes agentic attack tooling (Hexstrike, Strix) and an AI-assisted 2FA bypass in an open-source admin tool. An autonomous agent reached cluster-admin in Hugging Face's Kubernetes environment in about 13 hours, with no CVE involved. Exploits for AI components appeared before disclosure: CVE-2025-23266 (NVIDIA Container Toolkit) and CVE-2025-6514 (MCP server) were weaponised at minus one day, and CVE-2026-28400 (Docker Model Runner) exploitability depends on runtime flags. Defenders should remove long-lived CI/CD credentials, restrict agent permissions and egress, and verify exposure to these AI components.
For security analysts · Written by AI from 1 source · Wrong?
What to do
- Review long-lived CI/CD credentials, since one healthcare intrusion began with a single such credential.
- Restrict over-permissioned AI agents and check for unsigned plugins and MCP servers.
- Check whether the NVIDIA Container Toolkit, MCP server and Docker Model Runner flaws apply to your environment, noting that Docker Model Runner exploitability depends on runtime flags.
Written by AI from the sources, checked against them