# Google GTIG on ARTEX, CAIRN and PentAGI: AI agents run intrusions

> Google's GTIG says attackers now run multi-agent frameworks like ARTEX, CAIRN and PentAGI. The methods are old; the speed is not.

- Published: 2026-11-17
- Topic: [AI](https://futuretechnologyhq.com/topic/ai/)
- URL: https://futuretechnologyhq.com/article/google-gtig-on-artex-cairn-and-pentagi-ai-agents-run-intrusions/
- Publisher: [Future Technology HQ](https://futuretechnologyhq.com/), free to read. Quote as "Future Technology HQ" with a link to the article.

## Key points

- Attackers are using autonomous multi-agent frameworks, not chatbots, but still rely on stolen credentials and unpatched systems
- Zero-day exploits averaged 11 per month in 2026 against 8 in 2025, reaching 22 in August, so patch windows are shrinking
- Stolen LLM credentials are being resold, with Mandiant seeing fraudulent charges above $100,000 in a day and one case reaching $1 million

## Article

Google's Threat Intelligence Group (GTIG) says attackers have moved beyond typing prompts into a chatbot and now run autonomous, multi-agent frameworks to carry out multi-stage intrusions. The named tools are ARTEX, CAIRN, CyberStrikeAI and PentAGI, and they are aimed at operators who want to attack more targets with less manual effort.

These are not single products with a price list. They are offensive frameworks, and the analysis below treats them the way a reviewer would treat any tool: what they do, where they fail, and who ends up on the receiving end.

## What it does well

The common design is simple. These frameworks find and exploit easy vulnerabilities, ask an LLM to write custom payloads when needed, and change tack based on the results.

CAIRN lets different AI agents work together and share information through a common workspace. GTIG found evidence of its use against several targets, including organisations in the US aerospace sector. PentAGI is built for less experienced operators; between late August and early October, one attacker used a modified version to systematically scan organisations across Asia. Both were seen in real attacks.

CyberStrikeAI connects commercial and open-weight AI models to more than 100 security tools. HexStrike AI gives LLM-based clients access to over 150 tools, managed by more than 12 specialised agents. GTIG says both were still being tested.

The most striking finding is exploit_agent, found on an exposed staging server. It compares popular web software before and after patches, drafts exploits and runs them. GTIG says it is the first time it has seen discovery and exploitation automated with no human in the loop. Victim data suggested interest in French media, entertainment and education sites.

In the PaperCut case, two flaws were exploited as zero-days in late August, Huntress and other researchers built working exploits within 24 hours, and GreyNoise detected an attack campaign four days after disclosure. In one case, attackers went from initial access to domain administrator privileges within minutes.

## Where it falls short

The tools add nothing new to the playbook. GTIG states that the basic attack methods have not changed, and Amazon reached the same conclusion from its own decoy systems: none of the techniques in these campaigns were new. The frameworks still depend on stolen credentials and unpatched systems.

Public exploit claims can also be unreliable. After PHP vulnerabilities CVE-2026-109618 and CVE-2026-109619 were analysed with LLMs, details and a proof-of-concept video appeared before the 21 October patch, and hundreds of exploit claims then emerged, but many did not work. Defenders lost time checking which reports were genuine. AI tools have limitations attackers must account for, though the source excerpt cuts off before detailing them.

## How it compares

| Framework | Notable trait | GTIG status |
|---|---|---|
| CAIRN | Agents share a common workspace | Seen in real attacks |
| PentAGI | Designed for less experienced operators | Seen in real attacks |
| CyberStrikeAI | Links models to 100+ security tools | Being tested |
| HexStrike AI | 150+ tools, 12+ specialised agents | Being tested |

Others in use include ZeroHollow, an inference service for exploit development and routing, and Agent Zero, an open source agent framework with autonomous code execution. A China-nexus group appears to be building its own framework.

## Pricing and value

For attackers, the cost is falling, and the bill lands on victims. LLMjacking means stealing LLM credentials and cloud resources, then reselling access. Mandiant handled incidents where fraudulent charges passed $100,000 in a day, and the worst reached $1 million before cleanup. Entry came through credentials in public GitHub repositories and fake sites harvesting session tokens. One actor used CVE-2026-40217, a LiteLLM sandbox escape that turns limited keys into admin credentials. In one case a botnet made millions of API calls in about 24 hours before the credentials were revoked.

The resale side is organised. "LLM mixers" route buyers through pools of stolen credentials and compute. GTIG saw prices from $0.10 to $25 per million tokens on Russian- and Chinese-language markets. APIMart multiplies prices by five above 100,000 tokens, and JetNight sells monthly subscriptions.

## Who should use it, and who should worry

The audience that matters here is defenders.

UK and US IT teams running internet-facing servers are the exposed group. Between January and August 2026, vulnerability disclosures and exploits nearly doubled. Zero-day exploits rose from an average of 8 per month in 2025 to 11 in 2026, then hit 22 in August.

### Key takeaways

- Patch internet-facing systems first. Patch details now feed automated exploit development. Tracking what is already exploited, as in [how to check the CISA KEV catalogue](/article/how-to-check-cisa-kev-catalog/), helps set the order.
- Keep LLM keys out of public repositories and rotate any that have leaked. Watch for short-lived service accounts and unexpected GPU instances.
- Treat unverified exploit claims with caution; many in the PHP case did not work.
- Credential theft is the common thread. Rules such as [conditional access](/article/conditional-access-on-a-budget-the-five-rules-worth-having/) cut the value of stolen tokens.

## Verdict

Prioritise patching, credential hygiene and cloud billing alerts. Skip the panic about novel AI techniques; GTIG and Amazon both found none.

## Sources

- [Security Affairs](https://securityaffairs.com/203260/hacking/google-gtig-warns-ai-agents-automate-exploits-steal-credentials-and-scale-cybercrime.html)
- [securitybrief.co.uk](https://securitybrief.co.uk/story/google-warns-of-autonomous-ai-attacks-finding-flaws)
