Google Spots Fully Autonomous AI Attacks For the First Time
Threat actors using AI do not need to find brand-new security flaws, says John Hultquist of Google Threat Intelligence Group. AI can now reverse-engineer security fixes automatically.
If earlier research from Google showed that attackers were starting to explore AI agents, this activity has now reached a major new level. Google says this marks a big change: cyberattacks can now happen at machine speed, with much less human involvement.
The Google Threat Intelligence Group (GTIG) published its AI Threat Tracker on 16 November 2026, working with Mandiant. The report is titled Agentic Threat Activity Crosses a Significant Threshold.
GTIG says this is the first time it has seen attackers fully automate the discovery of security weaknesses and the use of those weaknesses to break into systems. The report points to the growing use of AI agents and automated tools by attackers.
Attackers are now using automated systems made up of multiple AI agents. These include the widely discussed penetration-testing tool ARTEX and open-source tools such as CAIRN, CyberStrikeAI and PentAGI. They are also building their own systems to carry out several stages of an attack quickly.
Autonomous AI attacks: inside the exploit_agent framework
The report's biggest finding is that attackers are using tools that can find security weaknesses and exploit them without human help. The report describes how these tools work.
GTIG did more than observe this activity. It found a system called exploit_agent on a staging server that was open to the public. The system connects to several advanced and locally run large language models (LLMs), which are AI systems that understand and produce text.
The system checks code from before and after a security update is released. It looks for possible weaknesses, creates ways to exploit them and tests those methods without a person needing to take part.
The server also contained stolen data from victims and scripts designed to guess passwords, install webshells and remove webshells installed by other attackers.
Information collected from affected systems suggests that the operator targeted French websites in the media, entertainment and education sectors.
This is a step beyond GTIG's previous report, which described only partly automated scanning for security weaknesses and attempts to exploit them.
N-day vulnerabilities: why attackers target security patches
GTIG says attackers may find it easier to use AI to exploit n-day vulnerabilities than to find zero-day vulnerabilities. N-day vulnerabilities are security flaws that have already been made public and may already have a fix available.
This is because attackers can work from a limited set of information, such as different versions of the code and the notes released with software updates.
GTIG's data shows that the total number of publicly reported vulnerabilities and related exploits nearly doubled between January and August 2026.
The number of zero-day exploits increased only slightly compared with 2025. There were an average of 11 each month, compared with eight a month in 2025. However, the number reached 22 in August 2026.
GTIG says research from other organisations also supports its findings. This research shows that attackers are building fully automated systems that find and exploit security weaknesses.
One example involves security flaws in PaperCut print management servers. Attackers could join two weaknesses together to run their own code on a victim's system from a distance. This is known as remote code execution (RCE).
Security researchers and cybersecurity company Huntress created working ways to exploit the two linked PaperCut flaws within one day of their public disclosure.
GreyNoise reported that an attacker likely used LLMs to create and test exploits for these flaws just four days after they were disclosed. The attacker then targeted hundreds of organisations and gained control of a network. Within minutes, the attacker had increased that access to domain administrator level.
John Hultquist, Chief Analyst at the Google Threat Intelligence Group, says attackers are now using AI to work faster and attack more targets:
“Early on, much of the conversation around threat actor use of LLMs centred around finding brand-new, unknown flaws, or zero-days. 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.
This shows how the n-day problem is growing. Using this technology to work out how a security patch works is much easier than searching for a zero-day vulnerability.
John Hultquist, Chief Analyst at Google Threat Intelligence Group
“In our latest GTIG AI Threat Tracker, we highlight a case we found of an agentic framework reversing security patches autonomously without human intervention.
“This shows how the n-day problem is growing. Using this technology to work out how a security patch works is much easier than searching for a zero-day.”
“Businesses will soon face enormous pressure to install security updates faster. They may even have to make difficult choices to keep up.”
Multi-agent frameworks: automating attacks on live systems
GTIG and Mandiant have seen several groups of attackers using open-source AI penetration-testing tools to automate parts of real cyberattacks.
GTIG saw the CAIRN framework being used against targets that included organisations in the US aerospace sector.
It also tracked a customised version of PentAGI scanning organisations across Asia between late August and early October 2026.
Attackers are also testing CyberStrikeAI and HexStrike AI. A group with links to China also appears to be building its own framework, probably for spying.
For now, these tools are fast, but they are not very good at hiding what they do. When they run freely without human control, they leave traces that investigators can find.
GTIG expects people to become less involved as these tools improve. It also expects future attacks to become harder to detect and more advanced.
LLMJacking and influence operations: AI threats are growing
In 2026, Mandiant dealt with cases of LLMJacking. This is when attackers steal access details for AI models and cloud computing resources. They then use these services heavily, leaving the owners with large bills.
In some cases, the fake charges for just one day were more than US$100,000. In the most serious cases, the total cost of this misuse reached US$1m before the problem was fixed.
Attackers then sell access to these stolen accounts through illegal online shops called “LLM Mixers”. Prices range from US$0.10 to US$25 for every million tokens used by an AI model.
A company working as a commercial contractor used the Gemini API to build a ready-to-use influence platform. The platform was designed to manage more than 100,000 accounts.
A suspected North Korean group called MIDNIGHT NEPTUNE changed .claude/settings.json files at cryptocurrency organisations. This meant that starting Claude Code would trigger a hidden malicious program called the SOMBERMEME backdoor.
Google says it disabled the resources involved in these activities. Google DeepMind also used the findings to improve its systems for identifying threats and its AI models. Mandiant Consulting has published a guide to help organisations make their systems more secure.
The main findings from the GTIG report are:
GTIG and Mandiant published their AI Threat Tracker, titled “Agentic Threat Activity Crosses a Significant Threshold”, on 16 November 2026.
GTIG says exploit_agent is the first case it has found where attackers fully automated the discovery and use of security weaknesses without human involvement.
The total number of vulnerability disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits averaged 11 per month and reached 22 in August.
Researchers and Huntress created working exploits for two linked PaperCut flaws within one day of their disclosure. GreyNoise reported an attack against hundreds of organisations four days after the flaws were made public.
Mandiant dealt with LLMJacking cases where charges for a single day exceeded US$100,000 and the total misuse reached US$1m.
A commercial contractor used the Gemini API to build a platform designed to manage more than 100,000 accounts.
Google disabled the resources involved. Google DeepMind improved its threat-detection systems and models, and Mandiant Consulting published a security guide.
