Global – October 24, 2026 -- Agentic intrusions tripled in the second half of 2026 compared with the first half, according to CrowdStrike data, as threat actors hand hacking decisions and command execution to AI agents.
CrowdStrike's report, published on Wednesday, documents four threat actors. Two used sophisticated AI harnesses, the software that connects a large language model to external resources. Two used simpler setups. The firm said the main constraint on scaling such campaigns is "token budgets," not engineering complexity.
Spanish- and English-speaking actor breaches at least 15 Argentine healthcare organizations
The actor used a penetration-testing harness built on the open-source PentestGPT framework, modified with extra hacking scripts. One Anthropic Claude Code agent acted as supervisor and assigned tasks to a second agent. CrowdStrike observed the activity between mid-May and mid-August.
The agents pivoted between victim networks on their own. They also switched from reusing a stolen hashed password to using a Windows feature that silently opened a connection from a victim's domain controller to the attackers' infrastructure, which yielded five stolen passwords. The campaign also led to theft of medical records and other sensitive data from at least one target.
Chinese-speaking actor runs two harnesses against government and private targets
This actor orchestrated a penetration-testing framework with Claude Code across a wider range of attacks. One harness automated reconnaissance and assigned tasks to other agents. Claude Code could follow an elaborate standard operating procedure alone or delegate to specialist agents. The actor also built infrastructure to use OpenAI's ChatGPT and other services.
Targets included Brazilian and Colombian government agencies and Brazilian and American firms in technology, biotech and retail.
Two actors with basic setups still compromise government agencies
An Indonesian-speaking actor gave Claude Code more than 800 AI skills, largely collected from public GitHub repositories, and applied no custom orchestration. The actor sometimes had to step in to bypass safety guardrails. The campaign still compromised government agencies in Indonesia and Bangladesh, plus private companies CrowdStrike did not describe.
A Russian-speaking actor gave Claude Code a seven-part to-do list and "near-total autonomy" to choose intrusion methods. Its workflow sent activity notifications to three people associated with cybercrime activity. Victims included a U.S. help-desk platform, from which corporate client data was stolen, a Ukrainian government agency and software-as-a-service firms in several countries.
Monthly incident counts double in August and September 2026
CrowdStrike recorded more than twice as many likely agentic incidents in August 2026 as in July. September saw twice as many as August.
The firm expects agentic attacks to "almost certainly increase in volume and sophistication" over the next year. It said falling inference costs and larger model context windows will ease token budget constraints and let more actors scale operations.
CrowdStrike also expects agentic AI to help unskilled hackers match the capabilities of sophisticated cybercrime groups and nation-state teams over the next 12 months. Their campaigns will remain prone to amateur mistakes, but analysts said "their sheer scale" will still make the threat environment more dangerous.