Google Threat Intelligence Group and Mandiant said in a November 16 report that adversaries are using open source agentic frameworks to automate parts of intrusions and vulnerability exploitation, while noisy execution can aid detection.
What changed
Google Threat Intelligence Group and Mandiant said adversaries are moving beyond manual prompting to autonomous multi-agent offensive frameworks, including ARTEX, CAIRN, CyberStrikeAI, and PentAGI.
GTIG said it independently observed CAIRN and PentAGI deployed in the wild, and saw experimentation with CyberStrikeAI and HexStrike AI.
The frameworks can automate discrete phases and specific operational tasks during live intrusions, according to GTIG.
Technical details
GTIG described CAIRN as an open source multi-agent pentesting framework that uses an OODA loop and a shared blackboard to move from an initial parameter toward an operational objective.
CyberStrikeAI is described as connecting LLMs to over 100 security tools, YAML attack recipes, and custom scripts; HexStrike AI connects LLM clients to over 150 pre-installed offensive security tools.
GTIG said the frameworks currently trade stealth for velocity because unrestricted autonomous execution generates forensic traffic, relies on recognizable identifiers, and leaves code-level traces.
Vulnerability exploitation and LLMJacking
GTIG said it identified an exposed staging server hosting exploit_agent, a multi-agent vulnerability research and weaponization framework that compares code before and after patch announcements, drafts exploits, and automatically executes them without human intervention.
Victim telemetry on the server suggests the operator was interested in targeting French websites associated with media, entertainment, and education.
The report also cited PaperCut vulnerabilities CVE-2026-81578 and CVE-2026-82078, which the report said were exploited as zero-days that allow unauthenticated remote code execution when chained.
The report said researchers and Huntress generated working exploits within a day of disclosure, and Greynoise reported a campaign four days later.
The report said LLMJacking can generate significant usage and billing increases within hours, with single-day fraudulent charges exceeding $100,000 USD and cumulative abuse reaching $1 million USD in the most severe cases.
What Google says
Google said it uses safety classifiers, guardrails, threat intelligence, and containment protocols to counter misuse, and it introduced the Secure AI Framework.
Google said it deployed behavioral and volumetric rulepacks to detect ARTEX, and Mandiant released a hardening guide and architectural defense framework for LLM-enabled offensive tooling.
What to do
- Mandiant Consulting released a hardening guide and architectural defense framework designed to counter the speed and scale advantage of adversaries using LLM-enabled offensive tooling.
- Mandiant says hardening and mitigation guidance for defending against LLMJacking intrusions is available in its blog.
Key facts and where they come from
- GTIG said adversary tradecraft is shifting to autonomous multi-agent offensive frameworks.
Adversary tradecraft has continued to transition beyond manual prompting to deploying autonomous multi-agent offensive frameworks
- GTIG said it independently observed CAIRN and PentAGI deployed in the wild.
GTIG has independently observed open source tools CAIRN and PentAGI deployed in the wild
- GTIG said the frameworks automate discrete phases and specific operational tasks during live intrusions.
automate discrete phases and specific operational tasks during live intrusions
- GTIG said current frameworks trade stealth for velocity.
they currently trade stealth for velocity
- GTIG said exploit_agent automatically executes exploits without human intervention.
automatically executes them without human intervention
- The report said two PaperCut vulnerabilities were exploited as zero-days that allow unauthenticated remote code execution when chained.
exploited two vulnerabilities in PaperCut print management servers as zero-days that, when chained together, allow for unauthenticated remote code execution
- The report said single-day fraudulent LLMJacking charges exceeded $100,000 USD.
single-day fraudulent charges exceeded $100,000 USD
- The report said cumulative resource abuse reached $1 million USD in the most severe cases.
cumulative resource abuse reached $1 million USD before remediation
- Google said its SecOps team deployed rulepacks to detect ARTEX.
Google Security Operations (SecOps) has deployed specialized behavioral and volumetric rulepacks to detect use of ARTEX.
- Mandiant released a hardening guide and architectural defense framework.
Mandiant Consulting released a hardening guide and architectural defense framework
The open question: How should defenders balance detection of noisy agentic attacks with the risk of missing stealthier future tools?
Tell us what you think · Discuss it in the Threat Intel room
Reader poll: Should defenders prioritize detection of current noisy agentic attacks?
A show of hands from readers, not a scientific poll. Say why in the comments.
