Skip to content
Google Spots Fully Autonomous AI Attacks For the First Time

Google Spots Fully Autonomous AI Attacks For the First Time

Cybermagazine • November 17, 2026

If threat intelligence from Google revealed “nascent adversary interest in agentic capabilities”, such activities have now crossed a major threshold representing a “paradigm shift towards intrusions occurring at machine speed”, the company says.

In fact, the Google Threat Intelligence Group (GTIG) AI Threat Tracker, published on 16 November 2026 and developed with Mandiant, has it in its title: Agentic Threat Activity Crosses a Significant Threshold .

GTIG says it is the first time it has seen fully autonomous vulnerability discovery and exploitation and adversarial adoption of agentic frameworks are to blame.

This includes adoption of autonomous multi-agent offensive frameworks, like the now “highly publicised” penetration testing tool such as ARTEX, along with open source frameworks like CAIRN, CyberStrikeAI and PentAGI, along with custom-built harnesses to execute multi-stage intrusions fast.

Autonomous AI attacks: inside the exploit_agent framework

The biggest finding from the report is that threat actors are using fully automated vulnerability discovery and exploitation tools.

GTIG did not just observe the activity, they actually found the framework, which is dubbed exploit_agent, on an exposed staging server. It connects to several frontier and local large language models (LLMs).

The framework compares code before and after patch announcements, identifies potential flaws, drafts exploits and runs them without human intervention.

The server also held exfiltrated victim data and scripts for brute-forcing, installing webshells and removing other attackers’ webshells.

Victim telemetry suggests the operator targeted French media, entertainment and education websites.

This is an escalation from its report, where GTIG had detailed only partially automated vulnerability scanning and exploitation .

N-day vulnerabilities: why attackers target patches

GTIG says attackers may find patched n-day flaws, which are already publicly disclosed, easier to weaponise with LLMs than zero-days.

This is because they can focus on a “finite set of source materials” such as code versions and release notes.

GTIG’s data shows overall disclosures and exploits nearly doubled between January and August 2026.

Zero-day exploits grew “only marginally” from 2025, averaging at 11 a month against eight a month in 2025. Although it spiked to 22 in August 2026.

GTIG notes that open source research corroborates threat actor operationalisation of fully automated vulnerability exploitation pipelines.

The company points to the vulnerabilities in PaperCut print management servers as an example, where two flaws could be chained together to achieve remote code execution (RCE).

Security researchers and cybersecurity company Huntress generated working exploits for two chained PaperCut flaws within a day of disclosure.

GreyNoise reported that a threat actor likely used LLMs to build and test exploits for these flaws just four days after disclosure before hitting hundreds of organisations, expanding one foothold to domain administrator within minutes.

John Hultquist , Chief Analyst at the Google Threat Intelligence Group, says attackers are now using AI for speed and scale:

“Early on, much of the conversation around threat actor use of LLMs centred around finding brand-new, unknown flaws, or zero-days. The reality we’re facing today is threat actors using this tech to operate at speed and scale, and they don’t need zero-days to do it.

This underscores the growing n-day problem – it’s far easier to reverse-engineer a patch using this technology than it is to hunt for a zero-day John Hultquist, Chief Analyst at Google Threat Intelligence Group

This underscores the growing n-day problem – it’s far easier to reverse-engineer a patch using this technology than it is to hunt for a zero-day

“In our latest GTIG AI Threat Tracker, we highlight a case we found of an agentic framework reversing security patches autonomously without human intervention.

“This underscores the growing n-day problem – it’s far easier to reverse-engineer a patch using this technology than it is to hunt for a zero-day .

“Enterprises are going to be under enormous pressure to speed up their patch cycles in the near future, and may even have to make some hard choices to keep up.”

Multi-agent frameworks: automating live intrusions

GTIG and Mandiant have observed multiple threat actors using open-source agentic penetration-testing frameworks to automate phases of live intrusions.

GTIG saw CAIRN used against targets including the US aerospace sector.

It also tracked a customised PentAGI platform scanning organisations across Asia between late August and early October 2026.

Threat actors are also experimenting with CyberStrikeAI and HexStrike AI, while a group with a China nexus appears to be building its own framework for likely espionage use.

For now, speed comes at the cost of stealth, because unrestricted autonomous execution leaves forensic traces.

GTIG expects reduced human involvement and increased stealth and sophistication as the tools improve.

LLMJacking and influence operations: AI threats widen

Mandiant responded to LLMJacking cases in 2026, where attackers steal LLM credentials and cloud resources and generate significant usage running up the bill.

In some cases, single-day fraudulent charges exceeded US$100,000, with abuse reaching US$1m in the “severe cases” before remediation.

Access is then resold through “LLM Mixers”, illicit storefronts that price it from US$0.10 to US$25 per million tokens.

A commercial contractor used the Gemini API to develop a turnkey influence platform designed to manage more than 100,000 accounts.

A suspected North Korean cluster, MIDNIGHT NEPTUNE, altered .claude/settings.json files at cryptocurrency organisations so that launching Claude Code triggered the SOMBERMEME backdoor.

Google says it disabled the assets involved, while Google DeepMind has used the insights to strengthen its classifiers and models. Mandiant Consulting has also released a hardening guide for defenders.

GTIG and Mandiant published the AI Threat Tracker, “Agentic Threat Activity Crosses a Significant Threshold”, on 16 November 2026

GTIG says exploit_agent is the first case it has identified of threat actors fully automating vulnerability discovery and exploitation without a human in the loop

Overall disclosures and exploits nearly doubled between January and August 2026, while zero-day exploits averaged 11 a month and spiked to 22 in August

Researchers and Huntress generated working PaperCut exploits within a day of disclosure, and GreyNoise reported a campaign against hundreds of organisations four days after disclosure

Mandiant responded to LLMJacking cases where single-day charges exceeded US$100,000 and cumulative abuse reached US$1m

A commercial contractor used the Gemini API to build a platform designed to manage more than 100,000 accounts

Google disabled the assets involved, Google DeepMind strengthened its classifiers and models, and Mandiant Consulting released a hardening guide

Key partners of Google Cloud

Accenture : A strategic partner helping Google Cloud deliver large-scale cloud migration, AI and digital transformation solutions to enterprise customers worldwide.

Accenture : A strategic partner helping Google Cloud deliver large-scale cloud migration, AI and digital transformation solutions to enterprise customers worldwide.

Deloitte : Works with Google Cloud to provide cloud modernisation, data analytics, cybersecurity and AI services to businesses across multiple industries.

Deloitte : Works with Google Cloud to provide cloud modernisation, data analytics, cybersecurity and AI services to businesses across multiple industries.

Capgemini : Partners with Google Cloud to accelerate digital transformation through cloud infrastructure, application modernisation and generative AI solutions.

Capgemini : Partners with Google Cloud to accelerate digital transformation through cloud infrastructure, application modernisation and generative AI solutions.

HCLTech: Supports Google Cloud customers with cloud migration, engineering, infrastructure management and enterprise technology services.

HCLTech: Supports Google Cloud customers with cloud migration, engineering, infrastructure management and enterprise technology services.

NVIDIA : Provides the advanced GPU infrastructure and AI computing technology that powers Google Cloud's AI and machine learning capabilities.

NVIDIA : Provides the advanced GPU infrastructure and AI computing technology that powers Google Cloud's AI and machine learning capabilities.

John Hultquist Chief Analyst

US Bank Outages: Why Critical Infrastructure Keeps Failing Cyber Security

Palo Alto Networks: 99% of Companies Have Ungoverned AI Technology & AI

FIOR Secures AI Agent Identity with NCSC-Tested Gateway Technology & AI

People Moves: November’s Cyber Executive Appointments Cyber Security