1-minute read
What a defender needs to know
One attacker used AI agents to breach an enterprise network and seize root credentials in under 10 hours, according to a Palo Alto Networks Unit 42 incident response report. The operation crossed a public-facing web service, code repositories, a secrets manager, CI/CD workflows and cloud AI infrastructure. It used more than 50 MITRE ATT&CK techniques. Hard branch protection stopped an attempted Terraform backdoor, and the report does not establish a widespread campaign.
Recognition depends on timing across systems. Search endpoints for structured Markdown files, Python caches and paired asset folders. Correlate bursty API requests, rapid 401/200 HTTP state shifts, parallel authentications and unexpected model usage with repository access, secrets-manager events, new build workflows and cloud-key activity. These observations are leads, because legitimate development and automation can produce similar artifacts.
Inventory model endpoints, API keys, MCP gateways and AI integrations, including the identities, repositories and pipelines allowed to reach them. If the pattern appears, revoke exposed credentials and OAuth sessions together, freeze affected CI/CD pipelines, isolate cloud accounts and inspect SSH keys, serverless functions, container restart policies and cloud identities for redundant persistence. Preserve identity, repository, build and cloud logs across at least the observed 10-hour window. Unit 42 does not provide a complete incident-specific recovery test, so each team must define the expected clean state for its own environment.