Google Threat Intelligence Group says some attackers are moving beyond asking LLMs for help and are using agents to carry out larger parts of an intrusion. Its September 8 report describes a second-quarter case in which attackers compromised a cloud resource, then planned, built and executed a mass credential-harvesting campaign in under six hours.
Google also tracked activity by UNC6780, a financially motivated group targeting open-source software supply chains and attempting to mislead coding assistants and LLM-based security scanners. In May, the company observed controllers of the ACRSTEALER malware seeking configuration files used by Cline and Continue. Those files can hold API keys and model-routing endpoints.
The report documents theft of AI platform accounts and developer credentials, attempts to access proprietary models and research, and misuse of compromised cloud infrastructure to run unauthorized AI workloads. Google says some coordinated attempts to extract capabilities from its models involved more than 100 million prompts.
These findings draw on Google's threat tracking, Mandiant investigations and platform defenses. They describe observed campaigns and tactics, not an estimate of how often AI agents are compromised.