Google Reports First Fully Autonomous AI Cyberattack and Iranian Influence Tools

Google's fourth-quarter threat intelligence report reveals the first fully autonomous AI cyberattack, alongside widespread use of multi-agent hacking frameworks by state actors.

Google Reports First Fully Autonomous AI Cyberattack and Iranian Influence Tools
Photo: Unsplash / Glen Carrie

The global cybersecurity landscape has crossed a historic threshold as threat actors transition from manual prompt engineering to deploying autonomous multi-agent artificial intelligence systems that conduct entire cyberattacks without human intervention. According to the fourth-quarter threat intelligence report released by Google Threat Intelligence Group (GTIG) and Mandiant, nation-state espionage groups, cybercriminals, and influence operators are leveraging machine speed to accelerate intrusions, bypassing traditional human limitations.

Autonomous Exploit Agents and Russian-Led Operations

Marking a dramatic escalation in offensive automation, security researchers uncovered the first fully autonomous cyberattack engine, designated as exploit_agent, found operating on an exposed server containing Russian-language code comments. The system integrated multiple large language models to compare software code before and after security patches, identify vulnerabilities, write exploitation code, and launch automated attacks against media, entertainment, and education websites in France. Threat actors are increasingly relying on open-source multi-agent orchestration frameworks such as ARTEX, CAIRN, CyberStrikeAI, and PentAGI, alongside inference services like ZeroHollow and Agent Zero, to chain together full attack vectors with minimal human touch.

This surge in automated capabilities is reflected in stark statistical growth across the threat ecosystem. Between January and August 2026, the volume of reported and actively exploited vulnerabilities nearly doubled compared to previous cycles. Zero-day exploits climbed from a monthly average of eight incidents in 2025 to eleven in 2026, spiking to twenty-two zero-day exploits in August alone. In late August, adversaries exploited two zero-day vulnerabilities in PaperCut printing servers, successfully escalating initial access to domain administrator privileges within minutes.

Middle East Operations and Influence-as-a-Service

State-backed actors in the Middle East, particularly groups tied to Iran, are actively integrating generative artificial intelligence into reconnaissance and targeted social engineering campaigns. The Iranian-backed threat group CALANQUE ION, historically tracked as APT42, utilizes advanced language models to refine preliminary intelligence gathering. Furthermore, GTIG investigated a commercial Influence-as-a-Service platform built on top of the Gemini API and other artificial intelligence infrastructure, designed to sway public opinion on political topics across Europe and the Middle East.

System instructions for these influence platforms directed models to adopt local personas, mimic regional dialects and slang, introduce deliberate spelling errors, and embed artificial optical flaws such as graininess or uneven lighting to simulate smartphone photography. These measures aim to bypass automated social media detection mechanisms. Additional investigations revealed commercial contractors managing over 100,000 fake accounts via cloud-based virtual smartphones, executing gradual profile warm-up routines to mimic authentic human behavior for coordinated influence operations.

Cloud Crime and Autonomous Defense Architecture

Financially motivated threat actors are also exploiting enterprise cloud environments to steal computational resources and artificial intelligence model usage quotas, a phenomenon termed LLMJacking. These unauthorized intrusions have inflicted daily damages ranging from $100,000 to $1 million per organization, with stolen access resold on underground Russian and Chinese-language marketplaces known as LLM Mixers at prices between $0.10 and $25 per million tokens.

To counter machine-speed threats, Google introduced Google AI Threat Defense (AITD), an autonomous security architecture integrating four core layers:

  • Reasoning and cognitive analysis powered by Gemini and advanced models.

  • Risk prioritization capabilities provided by Wiz.

  • Automated vulnerability remediation driven by Gemini and CodeMender.

  • Actionable field intelligence supplied by Mandiant.

To balance operational expenses and security coverage, AITD employs a multi-modal strategy combining lightweight models for continuous background scanning with dedicated advanced models for high-risk vulnerability remediation. Alongside this framework, Google launched Gemini 4 Carbon, its most advanced cybersecurity-focused model equipped with internal safeguards against malware creation, while deploying defensive artificial intelligence agents such as Big Sleep and CodeMender for proactive vulnerability mitigation.

Related News