1-minute read of a 5 min report

Cloud Intruders Automate Mass Credential Harvesting in Under Six Hours

By

1-minute read

What a defender needs to know

Google Threat Intelligence Group observed a Q2 2026 attacker compromise a cloud resource and complete an agent-enabled mass credential-harvesting campaign in under six hours. GTIG says adversaries are moving from individual prompts to multi-agent workflows that manage scanning, recover from errors, and reduce the pauses defenders depend on.

The report also describes UNC6780 abusing developer and AI tooling. Its DUSTMAKER credential stealer targets CI/CD OIDC tokens, hides malicious files in .claude/, .vscode/, and .cursor/, launches scripts such as setup.mjs, disguises pipeline work as Copilot Setup, and deletes workflow logs. Valid publisher identities and build attestations can preserve a compromised trust path.

Hunt for unexpected OIDC-token access, trusted-publisher changes, package releases outside normal maintainer workflows, missing CI/CD logs, unfamiliar hooks in AI-tool directories, and unauthorized AI compute. Correlate staged packages with process execution so potential exposure and executed code remain distinct. Treat unexplained workflow-history gaps as containment triggers.

Disable malicious principals and workflows before rotating cloud, repository, package-registry, CI/CD, and AI platform secrets. Preserve project configuration, runner telemetry, audit events, and cloud logs. Exercise the response with an unauthorized publisher event and AI-themed workflow task; the expected result is linked detection, identity disablement, and credential revocation within six hours.

Primary sources