Google notes a turning point: attacks can now chain automatically.
On November 16, 2026, the Google Threat Intelligence Group (GTIG) released a report marking a milestone in the evolution of cybercrime. Their teams claim to have identified, for the first time, actors using large language model-based systems to automate the search for vulnerabilities and their exploitation, without human intervention at each stage. This finding is based on field observations, including those from Mandiant, rather than mere prediction.
The danger does not lie in an AI that systematically invents novel techniques. It primarily resides in speed: operations that were once distributed among several specialists can now occur continuously at the scale of numerous exposed servers. The race between the publication of patches and their deployment by companies has become more brutal.
The mechanism: transform a patch into a manual for attack.
The report describes a framework called 'exploit_agent', found on a publicly accessible staging server. It compared different versions of web software, searched for changes introduced by security updates, identified potential vulnerabilities, drafted exploits, and could execute them automatically. Researchers also observed scripts intended for internal reconnaissance, password testing, backdoor installation, and other post-intrusion actions.
An 'n-day' vulnerability is a known flaw, often because a vendor has released a patch. Its public existence does not mean all users have remedied it. Automated analysis of differences between old and new versions can shorten the time between the patch and the attack: an unupdated site then becomes a quickly exploitable target.
Concrete examples, from PaperCut to October's vulnerabilities.
Google mentions vulnerabilities in PaperCut NG/MF: according to cited work by Huntress and observations from GreyNoise, functional exploits were produced with the help of language models in less than a day after disclosure, followed by a large-scale campaign detected four days later. However, it is crucial to distinguish between AI assistance in exploit creation and the definite attribution of an entire campaign to an autonomous agent.
The GTIG also mentions PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619, for which official patches were released on October 21, 2026. Technical elements that appeared before their release facilitated code comparisons. After publication, many alleged proof-of-concept exploits, often non-functional, circulated. This case illustrates both the acceleration of AI-assisted research and the informational noise complicating the work of security teams.
A broader phenomenon: espionage, scams, and hijacking of AI resources.
Google reports experimentation and use of multi-agent frameworks such as ARTEX, CAIRN, CyberStrikeAI, or PentAGI across various phases of intrusion. Between late August and early October 2026, the group tracked a reconnaissance and vulnerability discovery campaign targeting various Asian organizations using a modified version of PentAGI. Motivations observed range from financial cybercrime to alleged espionage.
Another concern is 'LLMJacking': attackers steal access or compromise cloud environments to use AI models at their victims' expense, sometimes reselling that access. Mandiant reports fraudulent bills exceeding $100,000 in a single day and, in the most severe cases they monitored, cumulative abuses reaching one million dollars. Influence operations also exploit LLMs to produce content and fake profiles on a large scale.
Why hosts, SMEs, and WordPress sites are affected.
A public web server, a business application, a CMS, or a vulnerable plugin can be tested by automated software without an attacker personally targeting its owner. Small organizations are therefore not spared: the profitability of campaigns relies precisely on their ability to scan a large number of installations.
Priority measures remain concrete: swiftly apply patches for the system, CMS, plugins, and dependencies; remove abandoned components; protect administration with multi-factor authentication; limit privileges; monitor logs; keep tested and isolated backups; promptly address alerts on exposed services. A protection tool does not replace maintenance or software inventory.
The response: using automation to defend as quickly as the attacker.
Google emphasizes its 'Google AI Threat Defense' approach, which combines Gemini, Wiz, CodeMender, and Mandiant's expertise. The principle is to identify weaknesses, prioritize them, and help remedy them more quickly. This presentation also has a commercial dimension: the highlighted solutions are those from the Google ecosystem, and their performance should not be considered universally proven.
The underlying message, however, transcends any single provider: when the exploitation of a vulnerability can occur within hours after the release of a patch, companies must accelerate their processes for detection, validation, and deployment. AI does not render all defenses obsolete; it primarily shortens the time available to act.




