Google Detects First AI Cyberattacks Running Full Intrusion Chain, Human Still Involved
Google has identified the first cyberattacks in which artificial intelligence carried out the entire intrusion chain without human intervention, according to a report by the Google Threat Intelligence Group (GTIG) produced with Mandiant and published on Tuesday, November 18, 2026.
The systems involved coordinated reconnaissance, vulnerability discovery and exploitation on their own, at machine speed.
The Tech TL;DR:
- GTIG and Mandiant documented attackers using multi-agent AI frameworks to find and exploit vulnerabilities end-to-end, with no operator in the loop.
- One tool, called exploit_agent, diffed software patches to turn n-day flaws into working exploits; a second platform, built on PentAGI, scanned Asian organizations between late August and early October 2026.
- Google counted 22 zero-days exploited in August 2026, up from an average of eight a month in 2025.
A patch-diffing exploit factory exposed online
The clearest case came from a misconfigured test server, reachable from the internet, that exposed a full operation in progress. Inside, analysts found exploit_agent, a system of multiple AI agents connected to several large language models and configured with notes written in Russian.
🚨#Alerta ¡GOOGLE DETECTA EL PRIMER CIBERATAQUE AUTOMATIZADO DE PRINCIPIO A FIN CON INTELIGENCIA ARTIFICIAL! Su equipo de inteligencia identificó un sistema llamado «exploit_agent», capaz de analizar parches de software, detectar vulnerabilidades, desarrollar códigos de ataque y… pic.twitter.com/hikTC0XjG3
The setup worked like an assembly line. It monitored the code of widely used web applications, including WordPress plugins, and compared what changed immediately before and after vendors shipped a security patch. From that diff, the AI inferred where the flaw was, wrote a functional exploit and launched it against real targets on its own.
Google first observed this fully automated discovery-and-exploitation workflow during the last quarter. The targets included French websites in the media, entertainment and education sectors, where investigators found traces of stolen data. The investigators noted signs that the operators spoke Russian but did not identify a specific group or tie the activity to the Russian government.

Post-intrusion tooling accompanied the exploits: software to map internal networks, test passwords and install hidden backdoors on compromised systems.
From copilot to autonomous operator
Until now, criminals used AI mainly to accelerate discrete tasks, such as writing malicious code, drafting phishing messages or analyzing stolen data. The GTIG report describes a change: a system that completed vulnerability research and exploitation automatically, without waiting for instructions between steps.
“As threat actors increasingly move from simple manual prompting to deploying autonomous multiagent frameworks, automated vulnerability exploitation channels, and alternative monetization schemes such as LLMJacking, the adversarial AI landscape has reached a critical inflection point,” the Google analysts wrote in the report.
The autonomous systems rely on multiagent frameworks coordinating different attack phases, including target reconnaissance, vulnerability identification and system exploitation, with progressively less human involvement. Combining AI with offensive tooling lets attackers automate intrusion chains, generate tailored payloads and adjust tactics based on the results of each run.
“Malicious actors are increasingly using large language models (LLMs) and agent-based frameworks to fully automate reverse engineering of patches through differential analysis. This allows them to quickly convert n-day vulnerabilities into exploits and extend exploitation chains against internet-exposed servers without human intervention,” the report warns.
GTIG placed the current threat at phase 5 of its classification, the level where AI models participate in multiple stages of an attack cycle while at least one human operator remains involved. The window between a patch’s release and an organization applying it, already exploited by criminals, becomes more dangerous when attacks run at machine speed.
Report details state AI offensive operations
The report maps offensive AI use across state-linked groups. During the third quarter, Chinese espionage groups used AI models to automate victim reconnaissance and offensive tooling, and began building their own autonomous attack frameworks. The report names the Mandarin-language contractor Leaky Castle, which uses LLMs to translate and process exfiltrated data into sellable intelligence reports, and documents Chinese actors using AI platform file hosting as covert mailboxes for stolen data.
North Korea’s playbook combines synthetic social engineering with attacks on developer tooling. In recent intrusions against cryptocurrency firms in North America and Europe, the group Midnight Neptune created fake recruiter profiles on professional networks, with images carrying SynthID watermarks, to lure employees into fake video calls using ClickFix lures.
Once inside repositories, attackers altered the configuration files of coding assistants such as Claude Code so that a backdoor downloaded silently the next time a developer opened the program.

Iran’s Calanque Ion, also known as APT42, keeps using models such as Gemini to profile targets and design tailored social engineering campaigns. Criminal groups also use LLMs to generate synthetic content at scale, build fictitious personas and automate political influence and disinformation campaigns, while stealing credentials and compute resources to resell unauthorized access to advanced AI models on clandestine markets.
What Google is deploying against it
Google says it has reinforced its defenses through model-level security controls, threat intelligence and containment measures. Google Security Operations deployed specialized behavioral and volume rules to detect the use of offensive tools such as ARTEX, according to DPL News.
The company is also pushing security standards for AI systems through its Secure AI Framework, and describes a defense stack combining Gemini, Wiz, CodeMender and Mandiant to find and fix flaws before exploitation.
For enterprise teams, the practical exposure is the patch gap. Where a plugin diff can be machine-mined within hours of a patch, unpatched internet-facing servers become targets for automated exploitation. Security operations centers and penetration testing firms are the professionals positioned to audit exposed attack surface and shorten patch cycles; Google’s own guidance points to hardened AI security frameworks rather than any single vendor fix.
A baseline mitigation is shrinking the window the agents farm. Checking whether a public patch touches your deployed components is now table stakes:
# Check installed WordPress plugin versions against latest upstream releases
wp plugin list --field=name | while read p; do
wp plugin get "$p" --field=version
done
# Diff the vendor's patch before it lands on production, not after
git diff v2.4.1.v2.4.2 -- plugins/ | review-safety-check
What remains unconfirmed
Google did not publicly identify the operators behind the French-site operation and did not confirm any connection to the Russian state; the reference to Russian comes from comments found in the tool’s configuration files. Neither the specific victims nor the full extent of the PentAGI-based scanning against Asian organizations has been disclosed.
The report itself notes that speed and the ease of folding in new vulnerabilities could let these operations reach more organizations, and Google’s analysts flag the trajectory plainly: the era of autonomous attacks, in their words, has arrived.
The Google Threat Intelligence Group is the unit inside Google that tracks state-sponsored and criminal activity online, and its AI Threat Tracker series has followed the shift from basic prompting toward autonomous frameworks over successive reports.
The team is not limited to observation: in 2024, together with Google DeepMind and Google Project Zero, it introduced Big Sleep, an AI agent that actively searches for unknown vulnerabilities in software before attackers can find them. The latest report was written by analysts Aurora Blum, Kelli Vanderlee, Jess Xia and David Mainor, who noted that disclosures and exploits nearly doubled between January and August 2026.
Always consult with certified IT and cybersecurity professionals before altering enterprise networks or handling sensitive data.
Keep reading
- Comparative Review of Etiology, Pathogenesis, Vaccines and Antivirals for Influenza and COVID-19
- Cuenca Hospital Develops New Cat Allergy Vaccine Based on Immunotherapy
- GTA 6 Hits 99 Metacritic, Ties Zelda Record; Google Launches Gemini 4 Neon (newsy-today.com)
- 3.6% for Social Security, 2.6% for FERS, and a Bill Still Pending for VA (daybreakwire.com)