According to Google Threat Intelligence Group (GTIG), cyber threat actors increasingly use multi-agent AI frameworks into live operations.
According to a report published by Google Threat Intelligence Group (GTIG) on 16 November, in collaboration with Mandiant, the operational use of artificial intelligence by cyber threat actors is entering a new phase. Moving beyond the use of large language models (LLMs) for basic assistance, attackers are increasingly integrating AI agents and multi-agent frameworks into live operations, automating successive stages of cyber intrusions, vulnerability discovery, data analysis and influence campaigns.
The report identifies what Google describes as a significant threshold in agentic threat activity: One of the most consequential developments concerns the increasing use of AI to coordinate and execute complex operational tasks with limited human intervention, cyber threat actors adopting of autonomous offensive frameworks.
GTIG has observed threat actors using open-source tools including ARTEX, CyberStrikeAI, CAIRN and PentAGI to automate reconnaissance, identify vulnerabilities, generate exploitation payloads and adjust techniques according to execution feedback. In one case, a customised framework based on PentAGI was used to conduct systematic reconnaissance and vulnerability scanning against organisations in Asia.
GTIG also reports that overall vulnerability disclosures and exploitation activity nearly doubled between January and August 2026, while the monthly average of zero-day exploits increased from eight in 2025 to eleven in 2026, before reaching 22 in August. The growing use of AI therefore appears particularly relevant to the exploitation of known vulnerabilities, as attackers can shorten the time between patch publication and weaponisation.
The implications extend beyond conventional cyber intrusions. GTIG assesses with high confidence that information-operation actors have moved from experimenting with LLMs to integrating them into their standard activities. State-linked groups, commercial Influence-as-a-Service providers and financially motivated political-content networks are using AI to generate synthetic media, create localised online personas, customise messaging and automate publication across platforms.
One platform investigated in August 2026 combined the Gemini API with models from other providers to automate the lifecycle of influence campaigns, including persona creation, image and media generation, distribution and interactive posting. Another platform under development was designed to manage more than 100,000 accounts across major platforms. Separately, cybercriminals are exploiting compromised cloud accounts and AI-service credentials to obtain unauthorised access to computing resources and advanced models, a practice known as LLMJacking.
Mandiant reported incidents in which fraudulent usage charges exceeded US$100.000 in a single day and reached US$1 million cumulatively before remediation. Underground services known as LLM Mixers are further commercialising this access by reselling stolen credentials and computing capacity.
As documented ealier in 2025, PRC-nexus espionage actors have used LLMs for target research, reconnaissance, exploit development and analysis of stolen information. These examples illustrate how AI tools can become both instruments of attack and vulnerabilities within organisations’ own development environments. Nevertheless, Google notes that autonomous offensive frameworks currently tend to trade stealth for speed: their execution can generate distinctive network traffic and forensic traces, while LLM hallucinations may undermine operational reliability.
Why does it matter?
The findings identify a turning point in the economics and operational organisation of cyber threats. By automating labour-intensive tasks and connecting them into repeatable workflows, agentic systems can increase the volume, speed and reach of operations while lowering the technical threshold for less sophisticated actors. The principal challenge is not necessarily the emergence of entirely new attack techniques, but the accelerated discovery, combination and exploitation of existing weaknesses.
Defenders therefore face shrinking response windows, greater pressure on vulnerability management and the need to monitor the use of AI agents, credentials and development tools inside their own environments. At the same time, the report cautions against equating automation with complete autonomy: human operators continue to define objectives and select targets in many observed campaigns. The next stage will depend on whether increasingly capable agents can reduce their operational noise and reliance on human supervision. As AI becomes embedded in both offensive and defensive security workflows, securing the systems, identities and permissions through which agents operate will become an increasingly important dimension of cybersecurity governance.
