NewsPolicy
MITRE ATT&CK v20 adds direct and indirect prompt injection, AI-enabled exfiltration and two AI-era campaigns to the catalog
MITRE released ATT&CK v20 on October 13, SecurityWeek reported on October 19. Nine new Enterprise techniques and sub-techniques include Direct Prompt Injection (T1699), for instructions typed into a model's interface, and Indirect Prompt Injection (T1698), for instructions hidden in what the model reads; AI-Enabled Exfiltration (T1700), for attackers who use an organization's own agents to take data; AI Result Poisoning (T1608.007), for web content planted to steer AI search and assistant recommendations; and Impair Autonomous Defenses (T1685.007), for disrupting model-assisted malware analysis. Compromise Software Dependencies and Development Tools (T1195.001) now covers IDE and coding agent configuration files. Two new campaigns: C0066, the 2026 attacks on internet-exposed PLCs at more than 30 water and wastewater utilities in Minnesota that locked operators out, with attribution unconfirmed, and C0064, the AI-assisted breach of Mexico's government infrastructure that crossed from IT to OT and reached a municipal water utility. The release also adds application-specific password theft, a CAPTCHA bypass technique, nine ICS techniques and sub-techniques, 15 detection strategies and 38 analytics.
Why it matters
Prompt injection and agent-driven data theft now have ATT&CK IDs, so detection rules, threat reports and tabletop exercises can name them the way they name phishing. Map your agent controls to T1698, T1699 and T1700 and see which ones have no detection behind them.