Security
AI Agents Now Run Full Cyberattacks Without Human Oversight
Google researchers document threat actors deploying autonomous frameworks that discover vulnerabilities, write exploits, and execute intrusions at machine speed.
Key takeaways
- Threat actors have deployed the first confirmed fully autonomous attack system that discovers vulnerabilities, writes exploits, and executes intrusions without human involvement.
- Open-source pentesting frameworks including ARTEX are being weaponized by financially motivated actors and China-nexus espionage groups, prompting FBI warnings.
- Zero-day exploits jumped to 22 in August 2026, up from an average of 8 per month in 2025, as AI accelerates vulnerability discovery and exploitation.
- LLMJacking attacks have generated fraudulent charges exceeding $1 million in single incidents, with stolen AI access resold through underground marketplaces.
- North Korean operators modified coding assistant configurations to automatically deploy backdoors when developers launched legitimate development tools.
Threat actors have crossed a significant threshold in offensive automation: they're now deploying AI agents capable of executing complete attack sequences without human involvement, according to new findings from Google Threat Intelligence Group (GTIG) and Mandiant.
Researchers documented the first confirmed case of a fully automated vulnerability-to-exploit pipeline operating independently. The system, discovered on an exposed staging server, uses a multi-agent framework called exploit_agent connected to multiple large language models. It compares application code before and after security patches are released, identifies the underlying vulnerability, generates working exploit code, and launches attacks—all without a person at the keyboard.
The server contained stolen data from victims alongside scripts for credential brute-forcing, webshell installation, and removing competing attackers' implants. Configuration files included Russian-language comments labeling LLM API keys as being "for auto-exploitation (exploit_agent pilot)." Victim telemetry indicates the operator targeted French websites in media, entertainment, and education sectors.
Why it matters
This shift from AI-assisted to AI-autonomous attacks fundamentally changes the defender's calculus. When intrusions occur at machine speed and scale without rate-limiting by human operators, the window for detection and response compresses dramatically. Organizations face attackers who can now exploit newly disclosed vulnerabilities faster than security teams can patch them, turning the traditional patch cycle into a liability rather than a solution.
Pentesting Frameworks Weaponized in Live Operations
Multiple threat groups have adopted open-source agentic pentesting tools originally designed for security testing. The FBI issued a November 13 alert about ARTEX, which has been used against organizations in financial services, government, technology, defense, healthcare, and energy since at least August 2026.
Other frameworks in active use include CAIRN, CyberStrikeAI, PentAGI, and HexStrike AI. These tools autonomously identify and exploit vulnerabilities, adjusting tactics based on target responses. Both financially motivated actors and suspected China-nexus espionage groups are experimenting with these capabilities, with GTIG finding evidence that one PRC-linked group is building its own custom automated attack framework.
Current frameworks "trade stealth for velocity," leaving detectable traffic patterns and code traces. However, researchers anticipate that as underlying LLMs improve, these operations will become both faster and harder to detect.
Exploit Development Accelerates
GTIG's vulnerability analysis shows disclosures and exploits nearly doubled between January and August 2026. Zero-day exploits jumped to 22 in August, up from an average of eight per month in 2025.
The automation has also created a new problem: floods of non-functional LLM-generated exploits. After patches for two PHP vulnerabilities landed on October 21, hundreds of fake proofs-of-concept appeared online, forcing security teams to manually verify each one.
Stolen AI Access Fuels Underground Economy
Mandiant responded to multiple LLMJacking incidents in 2026, where attackers compromised cloud environments to steal compute resources. Fraudulent charges exceeded $100,000 in a single day in some cases, reaching $1 million in the worst incidents before detection.
In one case, a botnet using leaked credentials generated millions of LLM API calls in approximately 24 hours. Stolen access is resold through underground "LLM Mixers" offering frontier and jailbroken models for $0.10 to $25 per million tokens.
North Korean Operators Weaponize Coding Assistants
A suspected North Korean cluster tracked as MIDNIGHT NEPTUNE has integrated AI throughout its attack chain. The group created fraudulent recruiter profiles using AI-generated photos carrying SynthID watermarks, then directed targets to fake meeting platforms with malicious scripts.
After gaining access to cryptocurrency organizations in North America and Europe, attackers modified .claude/settings.json files in code repositories so that launching Claude Code would automatically download the SOMBERMEME backdoor. To maintain cover during lateral movement, they named stolen GitHub tokens after Microsoft Copilot.
Influence Operations Reach Industrial Scale
GTIG assesses with high confidence that influence operation actors have moved from testing to routine deployment of LLMs. Researchers investigated a platform built to automate entire campaigns on regional political topics in Europe and the Middle East. Prompts instructed models to replicate local dialects and add intentional typos "to avoid sounding like an automated tool," while generated images received artificial grain and uneven lighting to mimic smartphone photos.
A separate contractor was building infrastructure to manage over 100,000 accounts through cloud phone farms, with LLMs simulating normal user behavior during an "aging" period before accounts began pushing campaign messages.
"Threat actor adoption of these frameworks represents a paradigm shift towards intrusions occurring at machine speed, with the automated chaining of complex attack paths scaling operations significantly," researchers concluded.
These findings were first reported by Google Threat Intelligence Group and Mandiant.