Google Report Tracks the Shift Toward Agentic Adversarial AI

Summary

Google Threat Intelligence Group’s Q2 2026 report says advanced threat actors are moving from basic prompting toward agentic workflows and AI-enabled automation, reducing the time available for defenders to respond. In one incident, an actor used a compromised cloud resource to plan, build, and run a mass credential-harvesting campaign in less than six hours; another operation exposed more than 23,800 harvested secrets through an automated reconnaissance framework. The report links AI-assisted coding and expanding open-source AI infrastructure to growing supply-chain risk, including trojanized MCP servers, malicious packages, hidden workspace files, CI/CD token theft, and prompt injection designed to manipulate coding assistants or evade LLM security scanners. GTIG also describes attackers stealing proprietary models, prompts, source code, and research, hijacking cloud infrastructure for unauthorized AI workloads, and buying or harvesting credentials for services such as Gemini, Claude, and coding tools. Model-distillation campaigns against Google systems have included coordinated efforts exceeding 100 million prompts, although Google says it did not observe direct attacks on frontier models by tracked espionage or information-operation actors during the quarter. Threat groups used AI across reconnaissance, social engineering, malware development, vulnerability research, exploitation, post-exploitation, and influence operations, but the report says it has not observed fully autonomous exploitation against targets or breakthrough capabilities in live influence campaigns. Google says it responded by disabling projects and accounts, improving classifiers and model safeguards, deploying defenses intended to reduce unauthorized model extraction, and expanding its Google AI Threat Defense architecture. The report presents adversarial AI as a maturing operational risk while emphasizing that several observed activities remained experimental or constrained.