Skip to content
Google Detects First AI-Led Autonomous Cyberattack

Google Detects First AI-Led Autonomous Cyberattack

M.Rediff • November 18, 2026

Google found that vulnerability disclosures and their exploitation nearly doubled between January and August 2026.

Google has identified what it described as the first observed cyberattack in which an artificial intelligence-powered system independently discovered and exploited software vulnerabilities without human intervention, signalling a shift towards increasingly autonomous cyber operations.

The finding, detailed in a report by Google Threat Intelligence Group (GTIG) and Mandiant, highlights how threat actors are moving beyond using generative AI as an assistant to deploying multi-agent systems capable of planning and executing multiple stages of an intrusion.

These systems can automate reconnaissance, identify weaknesses, generate malicious code and adjust their tactics based on the results of actions.

Google identified what it described as the first observed cyberattack in which an AI-powered system independently discovered and exploited software vulnerabilities without human intervention.

AI-powered multi-agent frameworks can automate reconnaissance, identify software weaknesses, generate malicious code and execute cyberattacks.

Google found that vulnerability disclosures and exploitation nearly doubled between January and August 2026, with monthly zero-day exploits reaching 22 in August.

AI-assisted attackers are increasingly targeting organisations through espionage, social engineering, vulnerability exploitation and automated intrusion campaigns.

LLMJacking and other AI-related cybercrime incidents have caused fraudulent charges exceeding $100,000 in a single day, with cumulative losses reaching $1 million in some cases.

AI-Powered Autonomous Cyberattacks

Google discovered an exposed staging server hosting a multi-agent framework called exploit_agent, connected to several advanced, locally hosted large language models (LLMs).

The framework was designed to examine software updates, compare code released before and after security patches, identify vulnerabilities and develop exploits to take advantage of them.

Once a weakness was identified, the system could automatically execute the exploit.

Scripts found on the server supported activities following a successful intrusion, including internal network reconnaissance, brute-force attempts and installing webshells to maintain access to compro­mised systems.

Based on the server's activity, Google assessed that its operator appeared interested in French websites associated with the media, entertainment and education sectors.

The discovery represents a significant development in AI-enabled cybercrime, with machines increasingly capable of carrying out interconnected attack operations rather than simply assisting human operators with individual tasks.

Multi-Agent Systems Enable Cybercrime

Google also identified growing adoption of open-source, multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI, PentAGI and HexStrike AI.

These platforms coordinate specialised AI agents to discover relatively simple vulnerabilities, generate attack code and modify subsequent actions using real-time execution feedback.

CAIRN, for instance, follows an Observe, Orient, Decide, Act (OODA) approach, allowing agents to assess their environment, evaluate available information and determine their actions.

Google observed the framework in intrusion activity targeting organisations, including those in the US aerospace sector.

It also tracked a threat actor using a customised framework based on PentAGI between late August and early October 2026 to conduct reconnaissance and vulnerability scanning against organisations across Asia.

Suspected espionage groups, including actors with links to China, have also experimented with developing their own automated attack frameworks.

Google additionally identified the use of tools such as ZeroHollow, an inference service for exploit development and LLM routing, and Agent Zero, an open-source framework supporting autonomous code execution and dynamic tool generation.

Software Vulnerabilities Under Attack

A major concern is the shrinking time between the disclosure of a software vulnerability and its exploitation.

Attackers increasingly use AI to examine differences between software versions before and after security patches, a technique known as differential patch analysis.

Zero-Day Exploits Nearly Double

Google found that vulnerability disclosures and their exploitation nearly doubled between January and August 2026.

Average monthly zero-day exploits increased from eight in 2025 to 11 in 2026, before reaching 22 in August.

The report cited the exploitation of two vulnerabilities in PaperCut print management servers in August 2026.

Security researchers and cybersecurity firm Huntress developed working exploits within a day of disclosure.

A subsequent campaign tracked by GreyNoise involved attackers using AI-assisted methods to build and test exploits, identify targets and launch attacks against hundreds of organisations.

In one instance, an attacker moved from an initial foothold to domain administrator privileges within minutes.

AI-generated exploits, however, are not always effective.

Google noted that the lower barrier to developing malicious code has also produced a growing number of exploits that are claimed to work but fail in practice, creating additional verification work for security teams.

AI Fuels New Cybersecurity Risks

The technology is also expanding the reach of espionage and influence operations.

Google said groups linked to Russia, China and Iran are using LLMs for target research, social engineering, data analysis and generating synthetic content for different audiences.

Financially motivated attackers are exploiting another emerging opportunity through LLMJacking, which involves stealing cloud credentials to access AI models and computing resources without authorisation.

Stolen access is sometimes resold through underground proxy services and marketplaces.

Mandiant investigated incidents across financial services, retail, transport and technology sectors in 2026, finding that unauthorised AI usage generated fraudulent charges exceeding $100,000 in a single day, with cumulative losses reaching $1 million in some cases.

Google also highlighted emerging supply-chain risks, including attempts to manipulate autonomous coding assistant configurations to introduce backdoors during local execution.

The findings suggest that AI is compressing attack timelines while lowering the technical barriers to conducting sophisticated cyber operations.

Feature Presentation: Ashish Narsale/Rediff

Extracted Entities