Google's Threat Intelligence Group (GTIG) and Mandiant reported on November 16, 2026, that they had confirmed for the first time an attack that automates everything from vulnerability discovery to exploitation without human involvement. In a September 8 report, GTIG said it had not confirmed any case of a fully autonomous attack system being used against real-world targets, so its assessment shifted in a little over two months. According to GTIG, however, even attacks using AI agents still rely on the same basic tactics as before, such as targeting compromised credentials and unpatched systems. What has changed substantially is the speed and scale of attacks, which leaves defenders less time to respond. In an attack exploiting a vulnerability in the print management software PaperCut, an attacker who began operating just four days after the vulnerability was disclosed breached real targets in under four hours.
A fully automated attack, unconfirmed in September, is first confirmed in November
What GTIG identified was "exploit_agent," a multi-agent framework found on a staging server that was accessible from outside. The framework connects to multiple frontier models and local large language models (LLMs) and continuously compares code from widely used web applications before and after patches are released. From those differences it looks for candidate vulnerabilities, then writes and runs exploit code, carrying out the entire sequence without human involvement.
Besides data stolen from victims, the server held scripts for internal network reconnaissance, brute-force attacks, installing web shells, and removing web shells planted by other attackers. Communication records related to victims suggest the attacker was targeting websites in France's media, entertainment, and education sectors. The configuration files also contained comments in Russian. The analysis targets also included version 2.19.25 of the WordPress plugin "Spectra (Ultimate Addons for Gutenberg)," showing that the framework was examining how access control was implemented.
In its September 8 report, GTIG said it had not confirmed any real-world use of a fully autonomous attack pipeline. The November 16 report instead described this as the first confirmed case of vulnerability discovery and exploitation without human involvement. This, however, reflects a change in what GTIG has observed, not evidence that fully automated attacks emerged only after September. The observation periods differ between reports, and it is not known when exploit_agent began operating.
Meanwhile, GTIG assesses that agentic penetration-testing frameworks such as ARTEX currently raise attack speed but reduce stealth. Unconstrained autonomous execution leaves heavy traffic and distinctive code traces, making attacks easier to detect. GTIG predicts, though, that as AI models and attack tools grow more sophisticated, human involvement will decline and attacks will also become stealthier and more refined.
PaperCut attack began four days after disclosure, breaching 11 organizations in 26 seconds
GTIG cited the large-scale attack on PaperCut NG/MF, investigated by security firm GreyNoise, as an example of AI accelerating attacks. In late August 2026, an authentication bypass vulnerability in PaperCut, "CVE-2026-81578," and a vulnerability involving unsafe dynamic class loading, "CVE-2026-82078," were exploited as zero-days. Combining the two allows remote code execution without authentication. According to GTIG, researchers and the security firm Huntress used LLMs to build working exploits within a day of the vulnerabilities' disclosure. Real attackers, for their part, began developing and testing AI-assisted exploit code four days after disclosure.
According to GreyNoise's investigation, the attacker set up an empty working environment on August 31 and achieved remote code execution against real targets in under four hours. Roughly two hours later, the attacker obtained its first domain administrator privileges. When the large-scale attack began on September 1, it breached 11 organizations in just 26 seconds. The attack used hundreds of AI agents combining OpenAI's Codex harness and DeepSeek models, working in parallel on vulnerability research, writing and testing exploit code, selecting targets, and carrying out intrusions.
The damage GreyNoise confirmed spread across 48 countries, with 440 PaperCut instances operated by at least 395 organizations compromised. Domain administrator privileges were obtained at 12 of those organizations, with the time from initial intrusion to privilege acquisition ranging from a minimum of 5 minutes to a maximum of 144 minutes. In some cases, however, the attacker paused post-intrusion work for several days, so the attack did not proceed at the same speed against every target. Cases in which Cloudflare's Web Application Firewall (WAF) blocked the attack were also confirmed.
What matters in this case is that the attackers were able to develop exploit code quickly by using the content of the vulnerability fix as a clue. Once a patch is released, attackers can narrow their research to the limited information of the code before and after the fix. GTIG analyzes that, for attackers, quickly weaponizing already-disclosed n-day vulnerabilities with LLMs can sometimes be more efficient than broadly searching for unknown zero-day vulnerabilities. The fact that exploit_agent was analyzing WordPress plugins suggests that similar dangers extend to ordinary websites. On sites where patching is delayed, the publication of the fix itself becomes a clue for attackers.
Zero-day exploitation rose to 22 in August: what should we watch next?
GTIG's own statistics suggest that viewing the AI-driven shift solely as a shrinking window for patching deserves scrutiny. According to GTIG, the number of disclosed and exploited vulnerabilities roughly doubled between January and August 2026. Zero-day exploitation also rose from a monthly average of 8 in 2025 to a monthly average of 11 in 2026, reaching 22 in August. That is 2.75 times the 2025 monthly average (22 ÷ 8, our calculation).
GTIG cites these statistics as an example showing that LLMs are affecting not only the speed of vulnerability discovery and exploitation but also the types and severity of the flaws that get found. However, the report does not make clear whether August's 22 include the two PaperCut vulnerabilities. Nor can current statistics alone determine whether the August surge is a temporary fluctuation or the effect of AI-driven vulnerability discovery. If zero-day exploitation keeps rising, countermeasures for vulnerabilities with no available fix will become even more important, alongside rapid response after patches are released.
Defenders are also beginning to respond to AI-accelerated attacks. According to GTIG, Google Security Operations (SecOps) has deployed detection rules based on attack behavior and traffic volume to detect use of the agentic penetration-testing framework "ARTEX." Mandiant Consulting has also published system-hardening guidance and defensive architecture design principles to counter the speed and scale of LLM-powered attack tools.
Particularly noteworthy in this report is that both exploit_agent and the PaperCut attack relied on publicly available information. exploit_agent analyzed code differences before and after patching, and the PaperCut attack began four days after the vulnerabilities were disclosed. By contrast, the report presents no direct evidence linking the rise in zero-day exploitation to the automation of attacks by AI.
What will inform judgments going forward are GTIG's updates to its vulnerability statistics and new observations of automated attacks. If zero-day exploitation runs well above the 2026 monthly average of 11, at levels approaching August's 22, or if multiple fully automated attacks separate from exploit_agent are confirmed, the focus of countermeasures would broaden to include zero-day response. If such trends are not confirmed, the gap shown in this case, with exploit code built within a day of disclosure and real attacks starting four days later, will be an important reference for organizations setting patching deadlines.
