Security
Google Detects First Fully Autonomous AI Cyberattacks
Threat actors are now using AI frameworks to discover vulnerabilities, write exploits, and execute intrusions without human intervention.
Key takeaways
- Google documented the first fully autonomous cyberattack framework, exploit_agent, which discovers vulnerabilities and develops exploits without human intervention.
- Threat actors are focusing on recently patched n-day vulnerabilities rather than zero-days, using AI to reverse-engineer patches within days of disclosure.
- Overall vulnerability exploits nearly doubled between January and August 2026, with attackers compressing exploitation timelines from weeks to hours.
- LLMJacking attacks generated fraudulent charges exceeding $100,000 per day in some cases, with stolen AI access resold through underground marketplaces.
- Multiple open-source AI penetration frameworks including CAIRN and PentAGI are now being used in live intrusions against aerospace, cryptocurrency, and other sectors.
Autonomous AI systems are now conducting cyberattacks
Google's Threat Intelligence Group has documented the first cases of fully autonomous cyberattacks, marking what the company calls a "paradigm shift towards intrusions occurring at machine speed." According to a report published November 16, 2026, threat actors are deploying AI frameworks that discover vulnerabilities, develop exploits, and execute multi-stage intrusions entirely without human oversight.
The most significant finding involves a framework called exploit_agent, which Google discovered on an exposed staging server. This system connects to multiple large language models and operates by comparing code before and after security patches are released. It identifies potential vulnerabilities, drafts working exploits, and deploys them autonomously. The same server contained stolen victim data and scripts for brute-force attacks, webshell installation, and removing competing attackers' access. Telemetry indicates the operator targeted French media, entertainment, and educational institutions.
Why it matters
This represents a fundamental acceleration in the cyber threat landscape. Organizations now face attackers operating at computational speed rather than human speed, compressing the window between patch disclosure and exploitation from weeks to days or even hours. The shift forces enterprises to dramatically accelerate their security response cycles or accept that adversaries will consistently outpace their defenses.
N-day vulnerabilities become the primary target
Contrary to early speculation that AI would primarily help attackers find zero-day vulnerabilities, Google's data shows threat actors are focusing on recently patched flaws. These "n-day" vulnerabilities prove easier to weaponize with AI because attackers can work from a finite set of source materials including patch notes and code differences.
Overall vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits grew only marginally, averaging 11 per month compared to eight monthly in 2025, though they spiked to 22 in August 2026.
The PaperCut print management server vulnerabilities illustrate this acceleration. Security researchers and cybersecurity firm Huntress generated working exploits within 24 hours of disclosure. Just four days later, GreyNoise reported a threat actor—likely using AI assistance—had deployed those exploits against hundreds of organizations, escalating from initial access to domain administrator privileges within minutes.
"The reality we're facing today is threat actors using this tech to operate at speed and scale, and they don't need zero-days to do it," said John Hultquist, Chief Analyst at Google Threat Intelligence Group. "Enterprises are going to be under enormous pressure to speed up their patch cycles in the near future."
Multiple frameworks now in active use
Google and its subsidiary Mandiant observed threat actors deploying several open-source penetration testing frameworks in live operations. These include CAIRN, used against U.S. aerospace targets, and a customized PentAGI platform that scanned organizations across Asia between late August and early October 2026. Actors are also experimenting with CyberStrikeAI and HexStrike AI, while a China-linked group appears to be developing its own proprietary framework.
Currently, autonomous execution trades stealth for speed, leaving forensic traces that defenders can detect. Google expects this limitation to diminish as the technology matures.
Broader AI-enabled threats emerge
Mandiant responded to multiple "LLMJacking" incidents in 2026, where attackers stole credentials for large language model services and generated fraudulent charges. Single-day costs exceeded $100,000 in some cases, with the most severe incidents reaching $1 million before remediation. Stolen access is resold through "LLM Mixers" at prices ranging from $0.10 to $25 per million tokens.
A commercial contractor used Google's Gemini API to build an influence operation platform designed to manage over 100,000 accounts. Additionally, a suspected North Korean group modified configuration files at cryptocurrency organizations to trigger backdoor malware when developers launched the Claude Code assistant.
Google disabled the compromised assets, while Google DeepMind strengthened its AI classifiers and models based on these findings. Mandiant Consulting released a hardening guide for organizations defending against these threats.
These details were first reported by Google Threat Intelligence Group in collaboration with Mandiant.