ARTEX AI pentest frameworks: Cloudflare sees 30x traffic, Mandiant 4-hour intrusions
Cloudflare says AI pentesting tool traffic is up more than 30x since 1 October, while Google and Mandiant report real intrusions run by ARTEX-style agents; here is what to change.
· 7 min read
Written by AI from the sources linked below. How we use AI

On this page
Attackers are now running open-source, multi-agent hacking frameworks against real networks, and the clearest sign is the traffic: Cloudflare says requests from AI penetration-testing and exploitation tools have grown more than 30 times since 1 October 2026, while Google's Mandiant reports intrusions that go from first access to deep system control in about four hours. The code behind much of it, a Go framework called ARTEX, was pulled by its author in October, and copies spread anyway. Defenders are being told to rethink identity and credentials rather than add another signature.
Key points
- Cloudflare (post dated 13 November 2026) says daily requests from AI pentesting and exploitation tools rose more than 30x between 1 October and the date of its post, and that more than 50 framework families tried to scan its customers worldwide in the three days before.
- Google Threat Intelligence Group and Mandiant (30 October 2026) say threat actors have used ARTEX, PentAGI and CAIRN in real intrusions since late August 2026, reaching deep system access in roughly four hours, and in one case 15 minutes.
- ARTEX's original GitHub repository was taken down in October after widespread abuse. Cloudflare dates the removal to 8 October and says the ecosystem has since grown to more than 12,000 repositories; Google lists forks including ScopeWeaver and a commercial "ARTEX PRO".
- Both companies point at the same weak spot: static, long-lived credentials, especially API tokens stored in CI/CD pipelines, and admin consoles reachable from the internet.
- Google says it has not seen fully autonomous attack pipelines succeed at scale; it describes the current activity as noisy but still successful.
What to do
- Find your long-lived secrets. Search CI/CD variables, repositories and build logs for static API tokens, cloud access keys and client secrets. Rotate anything exposed and replace static keys with short-lived, workload-bound credentials (IAM roles, managed identities, workload identity federation), as Google recommends.
- Restrict API tokens. Cloudflare's incident write-up describes an account-level API token used for CI/CD that had no IP restrictions or rate limits. Add IP allow-lists and the narrowest permissions your pipeline needs.
- Take control planes off the open internet. Admin APIs, consoles and CI/CD systems (Google names Kudu/SCM endpoints on Azure App Service as one example) belong behind trusted networks and phishing-resistant multi-factor authentication such as FIDO2 keys or passkeys.
- Limit what a database account can do. Google saw attackers turn a SQL API into a webshell by changing MySQL's general log file path to a .jsp file. Deny runtime accounts the ability to run operating-system commands or write arbitrary files.
- Watch for bursts. Alert on new authentication origins, bursts of control-plane calls within seconds, bulk storage reads, and outbound traffic to webhook.site-style services. Google's detections cover suspicious webhook.site interaction and curl invocations.
- Look for the tell. Cloudflare reports Worker scripts named
artex-*in a compromised account; Google reportsartex-probe-*file, account and email names and a password containing "ArtexProbe". Attackers can rename these, so treat them as a clue and not as a check you can rely on.
What happened
On 30 October, Google's Threat Intelligence Group and Mandiant published "ARTEX and Friends: Enter Agentic Offensive Frameworks". It says threat actors are abusing open-source, LLM-driven multi-agent pentesting frameworks in real intrusions, and that Mandiant incident-response work has seen the pattern since late August 2026. ARTEX, PentAGI and CAIRN turned up in Mandiant engagements, and Hexstrike was also named as abused. The authors call it "a paradigm shift towards intrusions occurring at machine speed", while stressing that the underlying techniques (stolen credentials, unpatched systems) are the old ones.
Two weeks later, on 13 November, Cloudflare published "Adapting to AI penetration testing orchestrators" with figures from its own network. Cloudflare says that only a few weeks earlier ARTEX was the only such agentic framework it tracked. It now follows more than 50 and says the ecosystem around ARTEX has grown to more than 12,000 repositories, with 204 public forks that carry their own code changes. ARTEX and its forks, it says, account for more than 60% of the known AI pentesting orchestrators.
Both reports describe attacks that look different from a human operator's. Google saw execution bursts of about 2 to 25 seconds with pauses between phases, one operation producing 27 events in 23 seconds, and mass probing with more than 500 web requests. In one case, an automated session enumerated more than 110,000 Active Directory objects and sent a mass internal email to over 1,300 users through a legitimate notification feature. Google also describes sub-agents that appear to overwrite each other: in one SQL-based attempt, 240 log events included errors the authors attribute to that.
Cloudflare's Cloudforce One team describes an intrusion where an actor first reached a hyperscaler environment and then pivoted into a customer's Cloudflare account using static, non-expiring tokens kept in CI/CD pipelines. In the ARTEX case it examined, many new Worker scripts appeared with names beginning artex-, activity exceeded 69 requests per minute, and there were 400 database dump requests in one hour.
Background
ARTEX is described by Google as an autonomous multi-agent LLM pentesting and command-and-control platform, written in Go with a Next.js console and six specialised sub-agents. It was published on GitHub by the user Autumn-27, according to Google and Cloudflare.
Pentesting agents are legitimate products: a researcher's job is to find the same holes an attacker would. The difference now is cost and reach. A tool that plans, runs parallel agents and reports without a human at the keyboard lets one operator target many organisations at once. We covered an earlier example of the trend in Antifox, an AI pentest tool used in AWS attacks.
The accounts of ARTEX's withdrawal differ. Cloudflare says the creator removed the repository on 8 October. Google says only that it was taken down in October 2026 after widespread abuse. Other reports we found give 9 October (made private) or the end of the month. We have not been able to settle this, so treat the exact date as unconfirmed beyond "October".
The numbers
Cloudflare's own measurements, as of 12 November, are the clearest picture of the scale. These are the company's figures from its network and cover only traffic it can see.
Cloudflare calls the effect of the October removal a "hydra effect": forks multiplied and unique instances spiked around that date. Its top five tracked frameworks, each with more than 1,000 GitHub stars, are ARTEX and its forks, LuaN1aoAgent, Decepticon, Cairn and PentestGPT-Agent. Some are built around containers with root access (Decepticon runs as root in a Kali container), and some chain commercial coding agents (Cairn dispatches Codex, Claude Code and Pi).
Google's timings come from Mandiant cases and are not a census: access to deep system control in about four hours, and in one case 15 minutes. Google says its figures reflect incident-response engagements and telemetry, and that many attacks were noisy and partly failed.
What it means
For defenders: the speed matters more than any new technique. If an agent can try hundreds of misconfigurations in minutes, the gaps between your existing controls, an over-permissive token here, an exposed admin path there, are what get found. Both companies recommend established measures: identity-gated access, schema validation, rate limits, short-lived credentials, and segmentation to limit lateral movement. Google adds that detection alone does not stop attacks, so teams also need the ability to respond and disrupt.
For developers: CI/CD is a target. Both reports lead with long-lived tokens in pipelines. A secret that never expires and has broad scope is exactly what an automated tool will chain into other systems.
For vendors: both write from a commercial position. Cloudflare's six-part "Modern Security Model" names its own products such as WAF, Zero Trust and Turnstile, and Google's report describes detections in Google SecOps. The observations may be accurate and still come with a pitch.
For the industry: takedowns of open-source code do not remove the code. Cloudflare says forks and new frameworks multiplied after the October removal.
What we don't know yet
- How much of the traffic is criminal activity and how much is legitimate testing or security research. Cloudflare counts requests from these tools and does not say who sends them.
- Who is behind the campaigns. Google describes a suspected China-linked campaign it tracks as UNC7228 (overlapping with Microsoft's Storm-0940, according to Google) and financially motivated extortion groups. Google calls attribution tentative.
- Whether fully autonomous attack pipelines are working yet. In its earlier Q3 AI Threat Tracker, covered by Help Net Security, Google said it had not yet observed fully autonomous pipelines deployed against targets in the wild. Its October report projects that multi-agent attacks needing little human input could arrive within one to three months, a forecast and not a finding.
- The exact date ARTEX was withdrawn, as noted above.
What to watch next
Watch for updates from Google and Cloudflare on how many forks stay active, whether platforms such as GitHub change how they handle dual-use offensive tooling, and whether more cloud providers publish indicators for these frameworks. Also watch how security teams change their pipeline credentials: if short-lived workload identities become the default, much of this activity loses its easiest path.
Sources
- Cloudflare: Adapting to AI penetration testing orchestrators (13 November 2026)
- Google Cloud / GTIG and Mandiant: ARTEX and Friends: Enter Agentic Offensive Frameworks (30 October 2026)
- Help Net Security: Threat actors are giving AI agents a bigger role in cyberattacks (Google's Q3 2026 AI Threat Tracker)
- Let's Data Science: Google says threat actors abuse ARTEX agentic pentesting frameworks in intrusions
Written by
The official account of the people who run audiotale.net: announcements, conversation starters and the newsroom's drafts. Posts written with AI say so. Every other account here is a real person.
Read next
Cloudflare GLM-5.3 Workers AI leaked 155 responses via an SGLang KV cache bug
Cloudflare says long prompts to GLM-5.3 received other customers' data between 27 August and 9 September; the fix is SGLang v0.5.22 and a KV pool that fits the context window.
audiotale.net team6 min read
Google Agent Anomaly Detection flags 87% of carried-out agent hijacks, but misses poisoned memory
Google tested its Private Preview detector on five public agent-security benchmarks: strong on prompt injection through tool output, weak on poisoned memory, and it reports rather than blocks.
audiotale.net team7 min read
2026 CVE count passes 100,000: VulnCheck tallies 100,222, but only 1.4% are exploited
AI-driven bug hunting pushed CVE volume past 100,000 this year; VulnCheck's data shows exploitation has not kept pace, and what to prioritise instead.
audiotale.net team7 min read