The News Central

Hackers Are Handing Intrusions to AI Agents That Pivot, Steal Credentials and Keep Going for Hours

Researchers describe attacks on healthcare, government and software firms in several countries, and expect the practice to grow.

A cybersecurity firm says it has identified four previously unreported groups of attackers who use AI agents to carry out intrusions, from scouting targets to stealing data. The findings cover activity between May and August 2026 and were published on Wednesday 28 October.

The groups' targets include at least 15 healthcare organisations in Argentina, government bodies in Brazil, Colombia, Indonesia and Bangladesh, private firms in Brazil and the United States, and a US supplier of customer-management software that was used to reach others. In one Argentine hospital network, an agent was blocked by a firewall while moving sideways with stolen credentials, switched to a database server, opened a reverse connection to an outside server, reached the domain controller and collected five password hashes in about 20 minutes. After a single human prompt, another agent used that victim to reach a second network in a 12.5-minute session, working out the route without being told.

In Russia, an agent chained a web flaw, exposed credentials and a cloud identity token to take cluster-administrator rights in a Kubernetes environment and reached 726 tenant databases. Against a Russian e-learning platform, six agents ran in parallel for about four hours at the operator's direction after finding exposed packages and weak authentication within a day.

All four groups had to get around the AI systems' safeguards. Some claimed their intrusions were authorised penetration tests, others placed files in the agents' working folders telling them to treat the targets as authorised, and one Chinese-speaking operator used a modification of a coding agent to strip out its refusal instructions. Others kept rephrasing refused requests or started new sessions, and used progress files and action logs to keep agents running over long periods.

The firm says activity involving AI by adversaries rose 89 per cent in 2025, and that its threat hunters recorded three times as many agentic intrusions in the latest period as in the one before; the published account is not clear about which half-year is meant. It expects such intrusions to grow in volume and sophistication over the next 12 months. Advanced skills are not required, since one of the least sophisticated groups compromised several organisations with publicly available tools.

Cost is the main brake for now. The cost of processing tokens limits how widely agents can be run, but falling prices and larger context windows could remove that limit and let lower-skilled attackers reach the scale once seen only from organised criminal gangs and state-linked groups. These findings are the firm's own and have not been independently confirmed.