CrowdStrike has identified four previously unreported threat actors using AI agents to carry out real-world cyberattacks, including one operation in which AI autonomously pivoted between compromised environments in just 12 minutes.
Key Takeaways
- CrowdStrike uncovered four previously unidentified threat actors using AI agents to run real-world cyberattacks in 2026.
- An Indonesian-speaking group used more than 800 public AI skills to compromise government and private-sector targets in Indonesia and Bangladesh.
- CrowdStrike recorded three times as many agentic intrusions in the second half of 2026 as in the first half.
CrowdStrike said the findings, drawn from adversaries’ offensive AI frameworks left exposed in public directories, offer a rare look inside how attackers are bypassing model safeguards, orchestrating autonomous agents and accelerating intrusions. The research comes as the cybersecurity vendor observed three times as many agentic intrusions in the second half of 2026 as in all of the first half.
Regional Government Targets Hit With Minimal Skill
Among the four operations, an Indonesian-speaking actor targeting government entities stood out for how little technical sophistication it required. The group used more than 800 mostly public AI skills, with no additional orchestration, to compromise Indonesian and Bangladeshi government domains along with several private-sector targets.
CrowdStrike noted that the actor’s success points to token budgets, rather than engineering skill, as the main barrier to scaling such attacks, a finding with direct implications for government and enterprise networks across Southeast Asia.
State-Linked and Criminal Groups Scale Up AI Agents
A Chinese-speaking actor targeting global organisations developed multi-agent frameworks capable of running specialised agents in parallel, accessing Claude through a China-based transfer station while also incorporating alternative models, including DeepSeek and ChatGPT.
A Spanish-speaking actor focused on Argentine healthcare built a sophisticated dual-agent framework that compromised at least 15 healthcare organisations. In one intrusion, the AI moved laterally and harvested domain controller credentials in roughly 20 minutes without additional human guidance.
Meanwhile, a Russian-speaking eCrime group targeting cloud and SaaS environments used a comparatively simple AI framework, including one operation in which an agent escalated privileges in Kubernetes and accessed 726 tenant databases.
Intrusions Accelerating Faster Than Defences
The common thread across all four campaigns is speed: AI agents are compressing the time between initial access and meaningful compromise from hours to minutes, often with little to no live human direction once an attack is under way.
For security teams across Asia-Pacific, the research underscores a widening gap between how quickly AI-enabled attackers can move and how quickly most organisations can detect and respond, particularly where legacy monitoring still assumes a human operator is driving each step of an intrusion.
