Stolen AI logins from 80,000 organisations found in malware logs

Credentials and session data for AI service accounts belonging to more than 80,000 corporate domains have turned up in infostealer logs, according to a report from the threat intelligence firm SOCRadar published on 28 September 2026.

Infostealers are malware that harvest saved passwords, browser cookies and session tokens from an infected machine and bundle them into logs that are then sold or traded. SOCRadar says those logs now routinely include logins for AI services, and that the exposure spans a large number of organisations rather than a single vendor or platform. The report does not attribute the theft to one breach at one provider; the credentials come from infected endpoints.

SOCRadar frames the consequences in two directions. The first is exposure of what employees put into AI tools. Anyone holding a valid credential or an unexpired session can potentially read stored conversations, which in a corporate setting may contain source code, internal documents, customer details or strategy material that never passed through an approved system. The report links this to shadow AI, the use of AI services by staff outside official procurement and monitoring, which makes the exposure harder for security teams to see.

The second is LLMjacking, which SOCRadar describes as the abuse of stolen AI accounts to run large language model workloads at the victim's expense. Because access to commercial models is metered and billed, a working login has direct resale value. SOCRadar says there is a growing market for stolen AI logins, and that the trade in them is now a distinct category within the wider infostealer economy.

The practical difficulty for defenders is that stolen session tokens can bypass password resets and, in some configurations, multi-factor authentication, because the token represents an already authenticated session. Rotating a password alone does not necessarily end an attacker's access if the session remains valid.

SOCRadar's report is aimed at helping organisations work out whether their own domains appear in circulating logs. The company said organisations can identify their exposure by checking stolen credential data against their domains, and positions the finding as part of a broader trend rather than a one-off incident.

The figure of 80,000-plus domains counts organisations with at least one exposed AI-related credential or session, not the number of individual accounts or the number of confirmed account takeovers. SOCRadar's report does not state how many of the stolen credentials were still valid at the time of analysis, nor does it name the AI providers involved.

For teams running AI agents, the exposure is not limited to chat histories. Accounts used to access model APIs are often the same accounts that authorise automated workloads, so a stolen key or session can mean both a confidentiality problem and an unexpected bill. SOCRadar's advice centres on discovery first: knowing which AI services staff actually use is a precondition for knowing which stolen logins matter.

The report was published on 28 September 2026. No vendor response has been reported alongside it.