Skip to content
NewA calmer, faster ZeroHour
Threat actorhighExploited in the wild

Google GTIG says AI agent frameworks now run multi-stage intrusions

Original title: Google GTIG Warns AI Agents Automate Exploits, Steal Credentials and Scale Cybercrime

Security Affairsby Pierluigi Paganini ()6 min read1
Part of a story covered by 13 sourcesGTIG says exploit_agent wrote and ran exploits with no human in the loopMerged summary & timeline
AI summary · x-ai/grok-4.7Summaries by Vani AI

CAIRN and PentAGI were used in real attacks, including against US aerospace, while exploit_agent built and ran exploits with no human in the loop.

Google’s Threat Intelligence Group says attackers have moved from chatbot prompts to autonomous frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI that chain scans, exploits and custom payloads. GTIG saw CAIRN and PentAGI in real intrusions and an exposed exploit_agent comparing patches and running exploits. The same report describes LLMjacking: stolen LLM keys and cloud compute resold through “LLM mixers,” with Mandiant citing charges up to $1 million.

Unconfirmed: A PRC-nexus group is only suspected of building its own framework, and many PHP exploit claims did not work.

Key facts
  • A modified PentAGI scanned organizations across Asia from late August to early October.
  • CyberStrikeAI and HexStrike AI were still being tested, GTIG said.
  • GTIG says zero-day exploits rose from about 8 a month in 2025 to 22 in August 2026.
  • PaperCut flaws were exploited as zero-days; domain admin followed initial access within minutes.
  • One actor used CVE-2026-40217, a LiteLLM sandbox escape, to turn limited keys into admin credentials.
Security assessment · AI, from the articleExploited in the wild

Impact: Faster n-day exploitation and theft of LLM credentials and compute

ProductAffected
PaperCut—
PHP—
LiteLLM—
Analyze this story with Vani AIImpact, who is affected and what defenders should do, in a new chat

Vulnerabilities mentioned

All
CVEVulnerabilityCVSSEPSSSignalsPublished
CVE-2026-109618
+1 related CVE
OOM longjmp memory flaw enabling RCE in PHP
9.2100%KEVPoC ×25
KEV
CVE-2026-40217
LiteLLM through 2026-04-08 allows remote attackers to execute arbitrary code via bytecode rewriting at the /guardrails/test_custom_code URI
8.817%PoC

From the original article

6 min read at securityaffairs.com

Google finds attackers using autonomous AI agents to automate exploits, steal AI access and scale intrusions, often sacrificing stealth for speed. Google’s Threat Intelligence Group (GTIG) says attackers have moved past typing prompts into a chatbot. They now run autonomous, multi-agent frameworks such as ARTEX, CAIRN, CyberStrikeAI and PentAGI to carry out multi-stage intrusions faster.

Security Affairs

By Pierluigi Paganini · securityaffairs.com

Read the full article at Security Affairs

A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.