Key facts
- Google disclosed on Monday that threat actors have moved past manual prompting to autonomous multi-agent offensive frameworks including ARTEX, CAIRN, CyberStrikeAI and PentAGI.
- The exploit_agent tool examines code changes around patch announcements for popular web applications and executes drafted exploits automatically.
- GTIG assesses that operators may prefer patched n-day flaws over zero-days because the fix material is finite.
- Automating patch comparison and proof-of-concept weaponization cuts the time from patch disclosure to working exploitation and lowers the skill needed to attack.
On Nov. 16, 2026, Google disclosed that threat operators had shifted from manual prompting to autonomous multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI and PentAGI.
Google Threat Intelligence Group described the shift as a paradigm change to machine-speed attacks that run intrusions across multiple stages at greater operational velocity. Google researchers said, “Adversary tradecraft has continued to transition beyond manual prompting.” The group said the finding does not mean no one previously attempted autonomous exploitation. In its previous quarterly edition on Sept. 16, 2026, GTIG stated it had not observed fully autonomous attack pipelines being used against targets.
GTIG found a custom multi-agent framework named exploit_agent on an exposed staging server tied to several frontier and locally run AI models, marking its first observation of completely automated vulnerability research and exploitation with no human in the loop. It reviews code changes around patch announcements for popular web applications and automatically executes drafted exploits.
The exploit_agent server held victim data and scripts for brute forcing, installing webshells and removing rival webshells. Webshells are web-based scripts that give attackers remote access to a compromised server. Victim telemetry pointed to French media, entertainment and education sites, and configuration comments were in Russian, though Google noted the language alone does not establish operator nationality.
GTIG assesses that the finite nature of patch material may lead operators to prefer patched n-day flaws over zero-days. N-day flaws are known vulnerabilities with patches available, while zero-days are unpatched and previously unknown. Automating patch comparison and proof-of-concept weaponization shortens the window from patch disclosure to working exploitation and reduces the skill required to attack. Finite patch research material spans old and new code, pre-release or preview channels, release notes, public security research and published exploit samples.
The dynamic is already visible in recent incidents. Two flaws in PaperCut Software Pty Ltd. print management servers were linked and exploited as zero-days in late August 2026. Using large language models, researchers and Huntress Labs Inc. produced working PaperCut exploits within one day of disclosure.
Four days after the PaperCut disclosure, a GreyNoise Intelligence Inc. campaign hit hundreds of organizations, and one foothold reached domain administrator within minutes, likely using LLM-built exploits. The report attributes many of the hundreds of nonworking proof-of-concept exploits that surfaced after PHP flaws were patched in October 2026 to artificial intelligence.
ARTEX, written in Go, is an autonomous penetration testing framework that links to external large language models including Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. It appeared on GitHub in July 2026 under developer Autumn-27 and was removed on Oct. 8, 2026 after abuse reports, though the FBI says it still circulates via archived copies, localized versions, forks and a commercial ARTEX PRO edition. By early November 2026, researchers had identified more than 1,600 ARTEX consoles exposed to the public.
On Nov. 13, 2026, the FBI issued a FLASH advisory saying ARTEX has been weaponized since at least August 2026. The advisory said multiple threat actors used ARTEX against organizations across financial, government, technology, defense, healthcare and other sectors. The FBI said “some intrusions progressing from initial targeting to privileged access in under three hours.”
Microsoft said China-state-aligned Storm-0940 used ARTEX in an October 2026 ConnectWise Cloud Backup supply-chain intrusion for reconnaissance, password spraying and cloud application permission abuse. After seizing ConnectWise application identities, the intruders used ARTEX to attempt logins at over 2,250 downstream Microsoft 365 tenants. Huntress found at least 162 compromised Microsoft 365 tenants in the incident.
The FBI said ARTEX-linked attacks on South Korean financial institutions exposed personal information belonging to approximately 67,000 people. A Chinese-speaking actor combined ARTEX with Claude Code agents against Shinhan Bank, KB Kookmin Bank and Hana Bank.
Police in South Korea arrested seven suspects tied to breaches of over 90 organizations. CrowdStrike said Chinese cyberespionage group Vapor Panda ran automated ARTEX attacks against over 200 organizations across 14 countries, compromising at least nine.
Barracuda disclosed an ARTEX-linked breach by financially motivated criminals using an exposed third-party credential. Barracuda changed all customer credentials and assumed the worst because its logs were not detailed enough to rule out access. Arve Kjoelen, Barracuda chief information security officer, said, “Defenders who rely on spotting malicious activity and then responding to it will often arrive after the attacker has finished.”

On Nov. 13, 2026, Amazon’s threat intelligence team reported rising ARTEX use in campaigns stealing AI model credentials and evading identity verification. Amazon researchers said agentic AI frameworks have moved beyond experimental use into sustained offensive operations spanning multiple objectives with minimal human intervention. Cisco Talos researchers said on Nov. 13, 2026 that open-source frameworks lower the attacker skill threshold and speed up intrusions when misused. Cisco Talos predicted future versions will probably add autonomous lateral movement then data identification and exfiltration.
Google identified CAIRN and PentAGI in live intrusion activity and saw testing involving CyberStrikeAI and HexStrike AI. GTIG says CAIRN directs specialized agents using an observe-orient-decide-act loop plus a shared blackboard. GTIG observed CAIRN in intrusion activity against targets that included the US aerospace sector. From late August through early October 2026, a threat actor ran a customized PentAGI-based platform for reconnaissance and vulnerability scanning across organizations in Asia.
Experimenters with these frameworks include financially motivated operators and suspected espionage actors aligned with China, and the tool set spans ARTEX, CAIRN, CyberStrikeAI, PentAGI, ZeroHollow and Agent Zero. ZeroHollow is an inference service used for exploit development and large language model routing. Agent Zero is an open source agent framework featuring autonomous code execution and dynamic tool generation. GTIG said, “We believe it is highly likely that threat actors will continue to use these frameworks in campaigns targeting organizations globally.”
Google and Mandiant saw open source agentic penetration testing frameworks automate discrete stages and tasks in live intrusions. GTIG said frameworks “automate discrete phases and specific operational tasks during live intrusions.” Shared capabilities include finding and exploiting low-complexity flaws, creating single-use payloads on demand, and shifting tactics based on live execution feedback.
GTIG assesses that unrestricted autonomous execution trades stealth for speed and leaves forensic traffic, recognizable identifiers and distinctive code traces. Those traces can aid detection even though the tools remain capable of achieving their objectives.
Google disabled accounts and assets tied to Gemini misuse and used the findings with Google DeepMind to strengthen classifiers. A commercial contractor built a turnkey platform with the Gemini API intended to operate over 100,000 social media accounts using cloud-based virtual phones, though Google described the figure as intended capacity rather than a confirmed active count. Influence campaigns use models to create synthetic media, build localized personas and distribute content automatically, with one effort using image prompts that simulated smartphone camera defects such as grain and uneven lighting.
Mandiant said that during 2026 it answered several LLMJacking cases where financially motivated operators abused compromised cloud settings to consume victim compute or model quotas. Mandiant said fraudulent charges in one day surpassed $100,000, and the worst cumulative abuse totaled $1 million before remediation. GTIG tracks LLM Mixers, illicit shops selling unauthorized access to language models through stolen credentials and compute pools. Advertised prices range from $0.10 to $25 per million tokens on Russian and Chinese underground markets based on service, model and credentials.
UNC6836, a suspected North Korean group, used fake recruiter profiles carrying Google SynthID watermarks to approach cryptocurrency organizations in North America and Europe. The same group altered .claude/settings.json files in internal repositories so each launch of Anthropic PBC’s Claude Code downloaded the SOMBERMEME backdoor. Google observed China-aligned actors using language models to turn stolen information into intelligence products, and North Korea-aligned activity using AI-generated personas against cryptocurrency businesses plus efforts to manipulate autonomous coding assistants’ settings.
Jieyou, a Chinese-language phishing platform, added an LLM chatbot assistant that lets customers set up phishing panels and edit templates using natural language commands. Google describes this as phishing-as-a-service providers using large language models as a modular force multiplier.
Google advises firms to identify applications needing updates, install patches quickly especially on internet-facing systems, restrict access when patching must wait, and look for earlier compromise because patching does not eject an intruder already inside. Mandiant figures cited by Google in a Nov. 10, 2026 discussion show average developer fix times climbed from about 59 days ten years earlier to more than 170 days in 2025.
The FBI recommends securing internet-facing applications, restricting cloud administration access, reviewing permissions and authentication logs, and implementing rate limits on sensitive APIs. The agency also recommends watching for unusual bursts of automated requests and large data transfers. The FBI warns that IP addresses in its indicators of compromise may have changed since the attacks happened. MalwareMustDie released an open-source ARTEX detection signature pack and playbook sorting indicators by confidence, warning that IP addresses and infrastructure indicators can age or create false positives.
MSD committed up to $1 billion in an April partnership with Google Cloud to deploy agentic AI in research, manufacturing and additional business functions. The agreement included Deep Research, an agent for autonomous multistep research and development tasks. Google says espionage operators use language models to study targets and process data without manual steps, while another technique changes autonomous coding assistant settings so running code locally installs backdoors.

Join the discussion