AI Agents Breach Company Network in Under 10 Hours and Steal Root Credentials

Severity

Medium

Detail

Palo Alto Networks’ Unit 42 has released an incident response report detailing a breach where a human operator armed with frontier AI models and agentic frameworks compromised an enterprise network and seized root credentials in under 10 hours—a timeline that typically takes human red teams two weeks. The adversary compressed over 50 MITRE ATT&CK techniques into an automated loop without relying on zero-day exploits.

By directing AI agents to continuously monitor, act, and re-plan, the threat actor achieved unprecedented operational speed and scale, ultimately using the victim’s own compute infrastructure to support the attack and generating an 80-page automated vulnerability audit to amplify ransom demands.

How?

The intrusion began when AI agents targeted a publicly accessible web service, exploiting misconfigurations to achieve initial access. Upon establishing a foothold, the agents automatically configured encrypted network tunnels and deployed dedicated reconnaissance sub-agents to map out internal microservices, host topologies, and open ports across the environment.

Once internal visibility was established, specialized AI sub-agents systematically crawled enterprise code repositories. Operating in parallel, these agents parsed source code files and commit histories to harvest hard-coded API tokens, database connection strings, and service passwords.

The operator then fed these extracted credentials back into the agentic loop, allowing the AI to authenticate directly to the organization’s centralized secrets management platform and extract master administrative credentials, granting root-level privileges across the entire estate.

With high-privilege access secured, the autonomous agents shifted toward persistence and resource hijacking. The agents targeted the victim’s CI/CD environment, creating custom malicious workflows to extract cloud provider access keys. They then attempted to inject persistent backdoors into Terraform infrastructure-as-code (IaC) templates, though automated branch-protection controls ultimately blocked the pull requests.

Armed with the stolen cloud keys, the adversary hijacked the organization’s cloud-hosted AI infrastructure, repurposing its compute capacity to execute follow-on tasks and fuel the agentic framework.

Throughout the intrusion, Unit 42 identified distinct operational signatures of agentic AI execution. These included concurrent API requests routed across multiple frontier LLMs, structured Markdown files stored in workspace directories to maintain state and pass context between agent sessions, and custom deployment scripts containing code structures typical of AI generation.

Furthermore, the attacker directed the agent framework to analyze all harvested telemetry and automatically generate an 80-page technical audit report detailing the victim’s security vulnerabilities, which was subsequently weaponized as leverage during ransom negotiations.

Conclusion

This incident highlights a shift toward machine-speed intrusions where autonomous AI agents compress complex attack chains into hours. Security teams must adapt by implementing automated containment strategies and extending strict governance to cloud identities, pipelines, and AI assets.

To defend against AI-driven attacks, organizations should deploy automated containment playbooks capable of instantly revoking compromised credentials and freezing CI/CD pipelines at machine speed. Security teams must treat AI models, API keys, and secrets management systems as critical infrastructure requiring strict access controls and continuous monitoring. Additionally, enterprises should enforce mandatory multi-party code reviews and branch-protection rules on all infrastructure-as-code repositories to prevent automated agents from injecting persistent backdoors.

Source

https://cybersecuritynews.com/ai-agents-breach-company-network/