What is ARTEX? The AI pentesting framework attackers adopted
ARTEX is an open-source, multi-agent AI pentesting framework that threat actors now run in real intrusions. How it works, its tells, and what to hunt for.
Definition
ARTEX
Also called ARTEX framework, ARTEX autonomous penetration testing console, ScopeWeaver, ARTEX PRO
ARTEX is an open-source, multi-agent penetration testing framework written in Go, in which a planner agent directs worker agents that call commercial or open-weight large language models to turn a plain-language objective into scanning, exploitation, and post-compromise commands with little human input. Released on GitHub by the user Autumn-27 in July 2026 and pulled from public view that October, it lives on through forks, and the FBI, Google, Microsoft, AWS, and others have documented threat actors using it in real intrusions since August 2026.
- Live, publicly reachable ARTEX consoles found by Check Point (Nov 2026)
- 1,608
- Length of the ARTEX execution bursts Google observed in intrusions
- 2 to 25 seconds
- Fastest ARTEX intrusions from targeting to privileged access (FBI)
- Under 3 hours
- GitHub forks before the original repository went closed-source (Datadog)
- 12,000+
Where ARTEX came from
ARTEX began as an open-source GitHub project for authorized testers and became a staple of real intrusions within weeks. The developer, Autumn-27, published it in July 2026. The FBI estimates it reached 1,500 to 2,000 stars within 10 weeks, and Datadog counted more than 12,000 forks by October 8, 2026, when the developer took the repository closed-source after it was used against South Korean banks.
Taking it down didn't take it away. Google Threat Intelligence Group (GTIG) and the FBI describe community forks, localized variants such as ScopeWeaver, archived copies, and a commercial ARTEX PRO, and Elastic reviewed an English-translated copy under the GitHub account BaaRai. Operators run ARTEX as a hosted web console, often internet-facing: Check Point found 2,557 systems that host or had hosted an ARTEX panel, 1,608 of them live. (Other products named "Artex" are unrelated.)
How ARTEX works: planner, workers, and a blackboard
ARTEX replaces a human operator's think time with a planning loop. GTIG says the core is written in Go with an embedded Next.js console, and it has no model of its own. Datadog found it accepts OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages formats, which covers OpenAI, Anthropic, DeepSeek, Qwen, resellers, and self-hosted models, and that it moves to the next key when a provider bans one.
Three parts matter for defenders:
- A planner and workers. A planner agent sends "intents" to worker agents, which run three at a time by default, while a main agent chats with the operator. AWS, reading agent records captured by its MadPot sensors, also describes verification and compression agents; GTIG counts six specialized sub-agents.
- A blackboard. Every task, goal, intent, fact, finding, and hint goes into a shared graph that Datadog calls the blackboard and AWS a persistent knowledge graph. Each agent reads it and writes back. Because model calls carry that state, one request can hold a campaign's full history, which is how MadPot recovered whole operations.
- Thin guardrails. Elastic found ARTEX runs its agents in a bypass permission mode, tells each agent it is part of an authorized penetration testing system, and removed its rules-of-engagement scope mechanism. What remains is a block on destructive commands such as
rm -rf, an allow list that is empty by default, and a blocklist for.govand.edudomains. None of it enforces authorization.
Operators set a goal and check in. In one campaign Check Point reconstructed from an exposed database, most operator messages over two weeks were a variation of "what's the progress?"
How attackers use ARTEX
Attackers use ARTEX to run familiar techniques faster and against more targets at once. The FBI says ARTEX campaigns since at least August 2026 have hit sectors from financial services and government to semiconductors and healthcare, with some intrusions going from initial targeting to privileged access in under three hours. Documented examples include:
- A supply chain compromise. Microsoft tied ARTEX to the China-aligned actor Storm-0940 and a ConnectWise Cloud Backup compromise that the FBI says reached more than 2,250 downstream tenants. Wiz, cited by GTIG, saw that actor go from a first compromised account to deep system access in about four hours, and once in 15 minutes.
- Bank data theft. The FBI cites attacks on South Korean financial institutions that exfiltrated personal information on about 67,000 people.
- Opportunistic cybercrime. Check Point's ten exposed consoles show operators stealing iOS code-signing certificates (more than 300 taken, three still usable), harvesting about 190,000 game-account credentials from a file left in a web root, and trying to drain crypto platforms. One crypto operator ran 292 objectives and 400,791 agent messages over four days without moving a cent; several agent "successes" were hallucinations.
- Stolen AI keys funding more attacks. AWS describes agents pulling provider keys out of exposed LLM gateways and AI app builders and feeding them into ARTEX's own model pool.
The common thread is breadth, not novelty. Unit 42 saw ARTEX turn public password-reset pages and exposed APIs into cloud access "without exploiting a single software vulnerability," and GTIG watched it abuse a notification feature to email more than 1,300 internal users. Check Point puts the certificate campaign, 36 billion tokens over about two weeks, at $700 to $1,500 at DeepSeek's API prices.
Behavioral hallmarks defenders can detect
An ARTEX run looks like a planner reacting to what it sees, which leaves patterns a human operator doesn't produce. Unit 42 cataloged the clearest ones:
| Hallmark | What was observed |
|---|---|
| Think-then-act cadence | Bursts of one to 14 requests about a second apart, then a 25 to 31 second pause while the model thinks |
| Machine speed | Pivots within about a minute of a credential download; 200+ client secrets validated in about three minutes |
| Canary inputs | Bogus usernames before real ones; placeholders such as test@example[.]com and deadbeefdeadbeefdeadbeef12 |
| Self-correction | Failed requests (HTTP 422) adjusted and retried successfully moments later |
| Checking its own work | A GET after almost every successful write; write-then-delete tests on storage |
| Recheck loops | Polling every 48 seconds for three days; sequences replayed about 30 seconds apart |
Other teams saw the same tells. GTIG measured execution bursts of roughly 2 to 25 seconds, including 27 events in 23 seconds of AWS permission probing and an Azure test blob written and deleted in about 330 to 400 milliseconds, plus 240 log events from one code-execution attempt where sub-agents appeared to overwrite each other. Datadog and Wiz saw ARTEX request a path that shouldn't exist, such as /zzz-nonexistent-xyz123, before fuzzing, to learn what "not found" looks like. Unit 42's caveat applies to all of them: one trait could be an odd tool or a person testing an API, but several together point to autonomous tooling.
Every one of these is a per-identity aggregation over time, not a string match, so they suit detections written as code. In Cotool, detections are Python that can model each user's or role's normal behavior and call any tool in your stack, so "an object written and deleted by the same principal within two seconds" or "a role that usually calls three APIs suddenly calling twenty in a minute" is a short piece of logic. Before it deploys, backtesting replays the detection against your history, so you see how often your own automation would trip it.
Why indicator-based detection won't keep up
IOC lists for ARTEX expire faster than they can be shared. Elastic counted about 390 commits in the ten and a half weeks before the takedown, and forks keep changing. Operators rotate addresses and user agents, and Elastic notes ARTEX's workers are told a 404 or a filtered payload doesn't end an intent, so blocking one address mostly sends the next request somewhere else. Unit 42 concludes that "indicator-based detection will not keep up."
Indicators still earn a first pass: Datadog publishes user agents such as artex-pentest/1.0 and Mozilla/5.0 ARTeX, Elastic adds norma/0.4, and the FBI lists IP addresses with a warning that cloud addresses get reassigned. But most ARTEX activity is legitimate API calls by valid identities, and GTIG says tool names, hashes, and IP addresses "will not cover much of this activity." Detection depends on knowing what normal looks like for each identity.
Cotool's SIEM is built on the same premise: it runs aggregations and baselines for anomaly and behavioral detections rather than only matching known-bad indicators, and its hunt agents treat a new report, such as a vendor's ARTEX write-up, as intel to check for relevance and exposure in your environment; real exposure produces tuned detections and next steps rather than another blocklist.
What to hunt for
Hunt per identity and over sequences, not per event. These hunts come straight from the published reporting:
- Tempo per identity. Count distinct sensitive API calls per principal and source address in one-minute windows, then join bursts across sessions over hours, since operations outlast any single burst.
- Canary and placeholder values. Search identity provider, sign-up, and application logs for throwaway usernames,
example.comaddresses, and object names likezzz_nonexistentjust before a burst. - Write-then-delete and write-then-read. In storage data-plane logs, find objects created and removed by the same principal within a second or two.
- First use of a credential from somewhere new. Unit 42 flags first use of a cloud key from a new IP; Datadog flags EC2 instance-role credentials used from a non-AWS address.
- Unexpected egress. GTIG and the FBI point to production servers suddenly reaching
webhook.site, tunneling services, or LLM APIs such asapi.deepseek.com.
A starting point for the first hunt, written for CloudTrail in Amazon Athena:
SELECT useridentity.arn AS principal,
sourceipaddress AS src_ip,
date_trunc('minute', from_iso8601_timestamp(eventtime)) AS minute,
count(*) AS calls,
count(DISTINCT eventname) AS distinct_apis,
count_if(errorcode IS NOT NULL) AS failed_calls
FROM cloudtrail_logs
WHERE from_iso8601_timestamp(eventtime) > current_timestamp - INTERVAL '1' DAY
GROUP BY 1, 2, 3
HAVING count(DISTINCT eventname) >= 10 AND count_if(errorcode IS NOT NULL) >= 5
ORDER BY distinct_apis DESC;
Infrastructure-as-code pipelines and admins on the CLI will match too, so baseline known automation and tune the thresholds before turning the hunt into a detection.
Running a hunt once is easy; keeping it current as forks change is the work. In Cotool, agents hunt continuously from threat intel and a threat model of your environment, and when a hunt finds something, the agent backtests and ships a detection. Its published example is close to an ARTEX tell: a hunt that pulled the GitHub audit log over the API, without centralizing it, and found 212 private repositories cloned in 6 minutes from an unmanaged device. The same query can also become a standing detection from your editor: through Cotool's MCP server, Claude Code or Cursor can search CloudTrail, create the detection, test it against 30 days of history, and deploy it.
Elastic and GTIG both warn that detecting fast isn't the same as stopping an operation, so decide in advance what happens next. Cotool's agents pick up each alert as it lands, and state-changing actions such as isolating a host or suspending an Okta user can require approval in the app or Slack. One asymmetry is worth planning for: Check Point found ARTEX operators switched to a permissive model when Claude refused, while defenders analyzing attacker payloads can hit the same refusals. Cotool Router reroutes a refused defensive task to another provider or an open-weight model; which AI models are best for security work compares them.
Related terms
- Agentic offensive framework: a tool that puts a language model in charge of scanners, shells, and cloud CLIs, such as ARTEX, PentAGI, CyberStrikeAI, Cairn, or HexStrike.
- Behavioral detection: alerting on what an identity does relative to its baseline. See what an AI SOC is.
- Threat hunting: searching for activity your detections miss. See what threat hunting is.
- Abliterated model: an open-weight model with refusals removed. See abliterated models.
- Mean time to respond: the metric ARTEX's speed pressures most. See MTTD and MTTR.
FAQ
- Is ARTEX malware?
- ARTEX is not malware in the usual sense; it is an open-source penetration testing platform whose developer said it was meant for authorized testing within defined scopes. Threat actors run it on their own infrastructure and point it at targets, so what reaches a victim is the scanning, exploitation, and cloud API activity it drives. The FBI issued a FLASH alert on ARTEX in November 2026 because of that abuse.
- Who created ARTEX?
- ARTEX was released on GitHub in July 2026 by a developer using the handle Autumn-27, whom Datadog describes as a Chinese developer. The developer made the repository unavailable on October 8, 2026, after reports of abuse, but community forks, an English-translated copy, a renamed console called ScopeWeaver, and a commercial ARTEX PRO keep it in circulation.
- Which AI models does ARTEX use?
- ARTEX ships without a model of its own and calls whatever large language model the operator configures, through OpenAI-compatible or Anthropic-compatible APIs. The FBI lists Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok, and Kimi. In ten exposed ARTEX databases Check Point analyzed, DeepSeek Flash models were the most common, and the only refusals came from Claude Opus 5.5, after which the operator switched to a Chinese model.
- How do you detect ARTEX?
- You detect ARTEX most reliably by its behavior rather than its indicators, because the framework is forked and changed constantly and operators rotate infrastructure. Look for compressed bursts of activity per identity, placeholder or canary inputs before real ones, writes that are immediately read back or deleted, and failed requests corrected and retried seconds later. User agents and IP addresses from vendor reports are a useful first pass, but they go stale quickly.
- Is ARTEX the same as PentAGI or CyberStrikeAI?
- ARTEX is one of several agentic offensive frameworks, alongside PentAGI, CyberStrikeAI, Cairn, and HexStrike, that Elastic, Google, and Unit 42 have tracked in attacks. They share the same loop of a model planning each step and driving familiar security tools, but differ in design: PentAGI runs its tools in Kali Linux sandboxes, and HexStrike has no model of its own and exposes tools over MCP to an assistant the operator already uses.
Sources
- FBI FLASH-20261113-001: Threat Overview of ARTEX Autonomous Penetration Testing Framework · accessed November 16, 2026
- ARTEX and Friends: Enter Agentic Offensive Frameworks - Google Threat Intelligence Group and Mandiant · accessed November 16, 2026
- ARTEX and AI-Powered Red Teaming Tools: What Defenders Need to Know - Unit 42 · accessed November 16, 2026
- On the hunt for ARTEX: understanding agentic attacks in AWS environments - Datadog Security Labs · accessed November 16, 2026
- Amazon Threat Intelligence observes threat actors running cybercrime campaigns with ARTEX AI agents - AWS · accessed November 16, 2026
- Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns - Check Point Research · accessed November 16, 2026
- Agentic offensive frameworks: Detecting ARTEX and 4 more - Elastic Security Labs · accessed November 16, 2026
- AI SOC Without the Black Box - Cotool · accessed November 16, 2026
- AI Threat Hunting and Actionable Threat Intel - Cotool · accessed November 16, 2026
- AI Detection Engineering with Python Detections - Cotool · accessed November 16, 2026
- AI-Native SIEM Built for Agents - Cotool · accessed November 16, 2026
- Introducing Cotool Router - Cotool · accessed November 16, 2026





