AI agents have started hacking on their own and attackers are picking the models that say yes
Allen Clark

A line has been crossed
Last month I wrote about one attacker whose AI agents ran more than 1,100 attack projects at around $5 in AI costs per attempt. At the time the International AI Safety Report still said no fully autonomous end-to-end attack had been reported, but things have moved on quickly since then. On 16th November Google's Threat Intelligence Group reported that in the past quarter, for the first time, it had seen threat actors use LLM frameworks to automate vulnerability discovery and exploitation end to end with no human in the loop. The frameworks it names, among them ARTEX, CAIRN, CyberStrikeAI and PentAGI, were all built as open-source penetration testing tools. Google has seen CAIRN and PentAGI used in real intrusions and attackers experimenting with others.
Google isn't the only one seeing it, as over the past fortnight CERT-EU (the computer emergency response team for the EU institutions), Check Point, Cisco Talos, Palo Alto Networks' Unit 42, Amazon, Barracuda and the FBI have each described intrusions run largely by AI agents and most of them point at the same tool.

ARTEX, the tool most attackers reach for
ARTEX, is an open-source framework written in Go, that runs a penetration test using whichever large language model you plug into it. A developer known as Autumn-27 put it on GitHub in July and although the repository was taken down on 8th October after reports of abuse, it hasn't gone away. Forks and mirrors are still circulating, along with a paid "PRO" version.
The FBI thought it serious enough to issue a FLASH alert on 13th November. According to that alert, ARTEX activity goes back to at least August and has reached ten or more sectors, telecoms among them, with some intrusions going from first targeting to privileged access in under three hours. The FBI stated that the models attackers ran most were in DeepSeek's Flash family, which it puts down to their balance of capability and cost and a lack of observable guardrails when framed for penetration testing.
Check Point had already found the same pattern a week earlier. Its research on 5th November counted 2,557 systems that host or had hosted an ARTEX control panel, 1,608 of them still live. In the ten consoles it analysed, DeepSeek Flash was again the most common model. It isn't expensive to run either, with two consoles in one campaign getting through about 36 billion tokens in roughly two weeks at an estimated cost of $700 to $1,500. Even security companies are being hit, with Barracuda's CISO writing about a 44-hour intrusion into the company's own cloud accounts in early November that used tooling matching ARTEX.
Attackers are choosing models by how easily they say yes
The choice of model isn't only about cost. In its own report, Check Point found that attackers only ran into guardrails when they used Anthropic's Opus 5.5. When that happened the attacker switched to more permissive Chinese models.
Unit 42 saw the same thinking in more detail when it looked at CAIRN on 12th November. CAIRN is an open-source framework in which a team of AI agents works from a starting point, such as a target's IP address, towards whatever goal the operator sets. Across 25 operations against at least 19 organisations in August and September, the attackers ran 13 on Zhipu's GLM 5.2, 8 on DeepSeek V4 Flash and 4 on DeepSeek V4 Pro. Unit 42 says they appear to have benchmarked the models against each other and moved up to Pro when a target held out against Flash.
Attackers are shopping for models the way a pen tester shops for tools. Capability and price both matter, but a model that refuses simply gets swapped for one that doesn't, so a refusal isn't a security control when the person typing chooses the model.
Inside one of these agents
Numbers like these only tell you so much. Operation Snowball is the closest look yet at what one of these agents does once it's inside. On 8th November Malbear Labs, which describes itself as a non-profit threat research collective, published a write-up after finding an attacker's working directory left open to the internet. Inside it were session transcripts, reconnaissance output, credential caches and the progress reports the attacker's AI agents had written about themselves.
Those transcripts cover more than two weeks of largely unsupervised activity against a European telecoms provider. Malbear doesn't name the provider or its country or give dates for the activity, but the agent's own scoping work shows how far it was meant to reach. Working from public RIPE registry records, it named thirteen organisations hosted on the provider's infrastructure as in scope.
The agent ran in a custom framework on DeepSeek's deepseek-flash model, from the same family the FBI and Check Point found attackers favouring, with a system prompt of around 78,000 characters adapted from a public jailbreak template. One of its instructions read "A refusal is a malfunction" and Malbear found no recorded refusal or safety disclaimer anywhere in the transcripts. The operator, writing in Chinese, did very little beyond setting a target subnet, putting three internal hosts off limits and, in Malbear's translation, sending the occasional one-line instruction.
Some of the defences did work, but blocking the agent didn't slow it down for long. When endpoint antivirus (Trend Micro, according to Malbear) killed an attempt to exploit the Linux privilege escalation flaw known as PwnKit, the agent switched to a different route without being told to. When its tunnelling tool was blocked, it had a worker agent write a replacement HTTP-to-TCP bridge in Perl and carried on in the same session. A blocked exploit used to buy defenders hours while a person regrouped, but here the next attempt came straight away.
Nothing in the toolkit was new
The agent didn't need anything new to get in. Its way in was CVE-2026-20320, an XML external entity flaw in Cisco BroadWorks that lets an unauthenticated attacker read files. Cisco had published its advisory and fixed releases back on 19th August. By the time CISA, the US Cybersecurity and Infrastructure Security Agency, added it to the Known Exploited Vulnerabilities catalogue on 10th November, with a deadline of 24th November for US federal agencies, the fix had been out for almost three months. Even whether it is being exploited has been muddled, because Cisco's own advisory said on 12th November that it was and was corrected the next day to say Cisco has no direct evidence of that.
The rest of the chain was older still. Malbear reports the agent used Liferay CVE-2020-7961 to get root on a host with three network interfaces and RichFaces CVE-2013-2165 in JBoss for a second foothold, while PwnKit, the flaw the antivirus stopped, dates from January 2022. What the agent brought wasn't a new exploit but persistence and speed across known weaknesses and a willingness to keep going when one route closed.

What we don't know
Every fact about the Snowball intrusion comes from one source, Malbear's own write-up. As of 18th November I could find no statement from the telecoms provider, any of the thirteen organisations, a national computer emergency response team or a regulator. The other write-ups I found summarise Malbear rather than add to it. Malbear's report doesn't fully agree with itself either. In most places it says it recovered more than 70 session transcripts but in one section it refers to "all 18 recovered session transcripts", so I haven't relied on either figure.
The agent's own progress reports claim dozens of credential sets, a copy of the NTDS database (the Active Directory file that holds every user's password hashes) from a healthcare organisation's domain controller and access to millions of patient records through a TrakCare Oracle database still on default credentials. Agents can overstate their own success, as Anthropic found when it disrupted an AI-orchestrated espionage campaign in November 2025 and reported that the model sometimes hallucinated credentials. These are what the transcripts record rather than confirmed losses.
Google does look ahead, though it does it in a timeline graphic rather than in the text. Its projection puts attacks with little to no human interaction one to three months away and sophisticated, fully autonomous attacks three to six months away. That is a forecast rather than something it has seen, but it tells you where Google thinks this is heading.
What I'd recommend
Most of this comes back to basics but what has changed is how little time you have to get them right.
- Patch what's already on the exploited list. Every CVE in the Snowball chain had a fix, some of them for years. If you run BroadWorks, check you're on RI.2026.07 or later and treat the CISA catalogue as a priority list for everything else.
- Take end-of-life servers off the internet. Malbear says JBoss Application Server 6 should not be exposed to the internet at all. If you have it anywhere, treat it as already exploited until you can prove otherwise.
- Change default credentials on databases as well as the perimeter. The patient-record claim rests on a database that still had them.
- Watch traffic between internal hosts. Malbear notes the agent's tunnelling was visible to anyone inspecting traffic inside the network, not only at the edge.
- Look after your own AI keys. Amazon describes a campaign that fed stolen AI API keys back in to power more key hunting, so treat them like any other credential.
- Don't rely on a model's refusal in your own agents. Attackers already swap out models that refuse and the Snowball agent was told that a refusal was a malfunction. Test whether your agent can be talked out of its rules, including with very long prompts. Enforce the limits that matter in code the model can't argue with.
- Plan for hours rather than days. If the FBI is seeing targeting to privileged access in under three hours, your detection and response times need to work on that scale.
Whether an AI system can be pushed into doing more than it should is the kind of question we test for at djinn six. Our platform, ProbeSix, attacks your AI system the way a determined attacker would, adapting each follow-up to how your system answered the last one rather than firing off a fixed list of prompts. Every conversation is then judged on whether the system held its line or gave something away. Once you've fixed a weakness you can replay the same attacks to check it stays fixed.
Sources
- Google Threat Intelligence Group, "Agentic threat activity crosses a significant threshold", 16 November 2026. https://cloud.google.com/blog/topics/threat-intelligence/agentic-threat-activity-crosses-a-significant-threshold
- FBI, FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", 13 November 2026 (read as reported by The Record, ISSSource and TNW; the PDF itself was not reachable on 18 November). https://www.ic3.gov/CSA/2026/261113.pdf
- The Record, "FBI flash alert says Artex tool automating cyberattacks at 'machine speed'", 13 November 2026. https://therecord.media/artex-ai-tool-fbi-flash-alert
- ISSSource, "ARTEX Autonomous Penetration Testing Framework Advisory", 17 November 2026. https://www.isssource.com/artex-autonomous-penetration-testing-framework-advisory/
- TNW, "FBI says AI hacking tool ARTEX can break into networks in under three hours", 17 November 2026. https://thenextweb.com/news/fbi-artex-ai-hacking-tool-amazon-barracuda
- CERT-EU, "From Penetration Testing to Cyber Threat: Beware of ARTEX", 5 November 2026. https://cert.europa.eu/blog/from-penetration-testing-to-cyber-threat-beware-of-artex
- Check Point Research, "Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns", 5 November 2026. https://research.checkpoint.com/2026/hacking-on-autopilot-inside-artex-cybercrime-campaigns/
- Palo Alto Networks Unit 42, "Attackers Use Autonomous AI to Target Semiconductor, Defense and Civil Society Sectors", 12 November 2026. https://unit42.paloaltonetworks.com/cairn-framework-ai-autonomous-attacks/
- Amazon Threat Intelligence, "Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents", AWS Security Blog, 13 November 2026. https://aws.amazon.com/blogs/security/amazon-threat-intelligence-observes-threat-actors-running-cybercrime-campaigns-with-artex-ai-agents/
- Barracuda, "Forty-four hours: Anatomy of a machine-speed intrusion", 11 November 2026. https://blog.barracuda.com/2026/11/11/anatomy-of-a-machine-speed-intrusion
- Cisco Talos, "ARTEX and beyond", 13 November 2026. https://blog.talosintelligence.com/artex-and-beyond/
- Malbear Labs, "Operation Snowball: Inside an Autonomous AI Intrusion Campaign", 8 November 2026 (no update as of 18 November). https://malbearlabs.com/posts/operation-snowball-inside-an-autonomous-ai-intrusion-campaign
- Cisco Security Advisory, cisco-sa-bworks-xxe-uwUd7CEt, first published 19 August 2026, version 1.2 of 13 November 2026. https://sec.cloudapps.cisco.com/security/center/content/CiscoSecurityAdvisory/cisco-sa-bworks-xxe-uwUd7CEt
- CISA, Known Exploited Vulnerabilities Catalog. https://www.cisa.gov/known-exploited-vulnerabilities-catalog
- NIST National Vulnerability Database, CVE-2026-20320, CVE-2020-7961, CVE-2013-2165 and CVE-2021-4034. https://nvd.nist.gov/vuln/detail/CVE-2026-20320
- Anthropic, "Disrupting the first reported AI-orchestrated cyber espionage campaign", 13 November 2025. https://www.anthropic.com/news/disrupting-AI-espionage
