Runtime as featured inForbesRead the article

What Is ARTEX? The AI Pentesting Tool Hackers Use, and How Fintechs Can Defend

ARTEX is an open-source AI pentesting tool attackers use to break in at machine speed. What it is, the FBI warning, Chase and Visa, and how fintechs defend.

Gus Trigos
Co-founder and CEO, Runtime
Updated November 18, 2026 · 18 min read

ARTEX is an open-source AI pentesting framework that uses a team of LLM agents to break into networks with little human help. It was built for authorized security testing. Since at least August 2026, attackers have used it in real intrusions, some reaching privileged access in under three hours, and the FBI lists financial services first among the targeted sectors.

Key facts, as of November 18, 2026

  • What it is: an agent harness, not a model. A planner agent turns a goal into attack tasks that worker agents run.
  • Who made it: a developer known as Autumn-27, who released it on GitHub in July 2026.
  • Where it is now: the original repository was pulled on October 8, 2026, but forks, archived copies, and a paid ARTEX PRO edition remain.
  • The warning: FBI FLASH-20261113-001, published November 13, 2026.
  • Chase and Visa: neither incident this week has been linked to ARTEX.

Attacks just got cheaper. A tool built to help security teams find holes now lets a small crew run dozens of break-ins at once, each one probing the partner portal, mobile backend, or forgotten API nobody watches. When the attacker's labor costs almost nothing, the defender's scarce resource is time.

This is a guide to what ARTEX is, what the FBI said, what did and did not happen at Chase and Visa, and how a fintech can defend itself with the same kind of agents. We run our own security this way, and leading fintechs use Runtime for cybersecurity alongside their payments and fraud work.

What it covers. The threat in plain language, the FBI's recommended controls mapped to what a fintech should do this month, and a setup for recurring AI pentesting and a swarm of watch agents.

What it is not. An attack manual. We do not link to ARTEX code or its copies. Everything here is for testing systems you own, within a written scope. Every material improvement comes from a human questioning a premise. Plan for that.

What is ARTEX?

ARTEX is an autonomous, multi-agent penetration testing platform written in Go. You describe a goal in plain language. A planner agent turns it into attack tasks, worker agents run scanners and exploits, and the results feed the next wave. It has no model of its own: it plugs into commercial LLMs or cheap open-source ones, and the FBI says DeepSeek Flash models dominate in attacker-controlled instances.

In other words, ARTEX is an agent harness: the software around a model that gives it tools, memory, and a plan. That is the same kind of software defenders can use, pointed the other way.

ReleasedJuly 2026 on GitHub, by a developer called "Autumn-27", for authorized testing
License and stackOpen source, written in Go, with a web console
PulledOctober 8, 2026, after reports of abuse
Still availableForks, mirrors, archived copies, and a paid edition, ARTEX PRO
Exposed consolesMore than 1,600 found online in early November 2026
What it attacksIdentity and cloud control planes more than new software bugs; the FLASH cites no CVEs

Is ARTEX malware? No. It is a dual-use tool, like other pentesting frameworks, and it has no CVE. There is no patch for ARTEX because it is not a vulnerability. It is a faster, cheaper attacker. The defense is hardening and speed.

Who uses it? The FBI names criminal and state-aligned groups. One China-linked group, tracked by CrowdStrike as VAPOR PANDA, used it against more than 200 organizations in 14 countries in a campaign that began October 17, 2026.

What did the FBI warn about ARTEX?

FBI FLASH-20261113-001, "Threat Overview of ARTEX Autonomous Penetration Testing Framework", was published November 13, 2026 (TLP:CLEAR). The points that matter for a fintech:

  • Speed. Some intrusions went from first targeting to privileged access in under three hours.
  • Scale. Small groups can run many intrusions at once. The barrier to entry is gone.
  • Financial targets. Attacks on South Korean financial institutions exposed personal data of about 67,000 people.
  • Supply chain. A compromise of ConnectWise Cloud Backup led to attempted sign-ins at more than 2,250 downstream tenants.
  • Overlooked surfaces. ARTEX is best at partner and supplier portals, mobile backends, employee apps, legacy services, and third-party integrations.
  • A short detection window. The attacks are noisy enough to trip alerts, but fast enough that teams have little time to respond.

Did ARTEX hit Chase, Visa, or other banks?

Short answer: no evidence links ARTEX to Chase or Visa. Three things happened close together, and they are easy to blur. Here is what is confirmed as of November 18, 2026.

EventWhat happenedLinked to ARTEX?
Chase outage, November 16Logins, transfers, and payments failed across the US. Chase blamed a software bug in on-premises hardware and said there is no evidence it was cyber-related. Customers at Bank of America, Wells Fargo, Capital One, and other banks reported problems on transactions with Chase.No
n0n extortion claim, November 16 to 17A new ransomware group listed Visa and JPMorgan Chase as victims and threatened to leak data within 48 hours. Visa said the incident involved a third-party provider in Asia Pacific, that its own systems were not involved, and that there is no sign card numbers were affected. The JPMorgan claim is unverified.No evidence
South Korean banks, September to OctoberARTEX-linked attacks stole customer data from auxiliary systems.Yes, per the FBI

Three lessons hold even where the facts are still unclear:

Your vendors are your attack surface. Visa's incident sat with a provider, and the ConnectWise campaign reached thousands of tenants through one vendor. ARTEX maps those trust links for the attacker.

Outages bring phishing. After the Chase outage, security experts warned customers about scam emails that use the disruption as a reason to "click now". Expect your brand to be impersonated the next time your app is down.

Silence is not a status. Customers and partner banks ask within hours. Being able to say what you checked, and when, is part of the defense.

How do attackers use ARTEX?

The FLASH maps the observed activity to MITRE ATT&CK. In plain language, the attacker's agents:

StageWhat the agents do
Find the way inScan for forgotten apps, try leaked or weak passwords, and look for admin passwords left in website code
Break inExploit common web flaws (SQL injection, IDOR, unsigned SAML) or log in with a vendor's stolen credentials
Stay inPlant web shells, add keys to existing apps, and reset passwords to take over accounts
Get adminSteal service tickets, query cloud metadata, and forge admin tokens with keys that were never rotated
Take the dataRead databases, mailboxes, and storage in bulk, then send it out through throwaway web hooks

The signature is volume and rhythm: bursts of requests, many failed attempts with small changes, sequential lookups on login and search endpoints, and unexpected calls to LLM APIs from your servers.

Leaked secrets: the fastest way in

Most ARTEX intrusions did not need a new exploit. They needed a key someone left in the open. According to Microsoft, as cited by GitGuardian, one ARTEX agent went from searching GitHub for exposed secrets to reading Azure Key Vault data in five minutes. The FBI lists hardcoded admin passwords in front-end JavaScript, credentials in files, stolen VPN configurations, and exposed cloud credentials among the ways in.

Check these places, because an attacker's agent will:

Where secrets leakExample
Public repos and forksA .env file, API key, or database URL committed once and never rotated
Git historyA key deleted from the code but still in an old commit
Front-end bundles and mobile appsAn admin password or API key shipped in JavaScript or an app binary
Container images and build logsCredentials baked into an image layer or printed by a CI job
Infrastructure code and pipelinesCloud keys in Terraform state, deploy scripts, or pipeline variables
Personal and contractor reposYour company's keys in an employee's or vendor's public project

Deleting a leaked key from the code does not make it safe. Revoke and rotate it, then check the logs for anyone who used it first. Prefer short-lived credentials so a leak expires on its own.

How can fintechs defend against ARTEX?

Start with the FBI's own recommendations, ranked for a payments company.

Do this monthWhy it matters against ARTEX
Inventory every internet-facing app, including partner portals, mobile backends, and old servicesARTEX goes for what nobody watches
Put cloud admin, deploy, hypervisor, and backup consoles behind private accessThese are the fastest route to privileged access
Rate-limit business APIs and cap transactions, recipients, and records per sessionMachine-speed abuse needs volume
Alert on sequential lookups and unusual records per sessionIt is the clearest automated-attack pattern
Audit third-party app access and consent grants; search by app IDThe supply chain is how one breach becomes thousands
Rotate long-lived keys and scan code, pipelines, and images for secretsUnrotated keys let attackers forge admin tokens
Stop showing stack traces and SQL errors to the publicError messages help the attacker's agents
Preserve logs before you revoke or delete anythingYou will need them for the FBI, your bank, and your customers

Then close the gap that matters most: time. An attacker that reaches admin in three hours beats a quarterly pentest and a team that reads alerts in the morning. You need to find the holes first and respond in minutes.

The defense: agents against agents

The answer to an automated attacker is not more people reading logs. It is your own agents, working inside rules you set.

Pentest scopeSigned by CISO
In scopeapi.staging, partner portal, mobile backend
Out of scopeProduction payments, third-party hosts
CadenceWeekly, and on login or payments releases
Stop rulePause and page on any real customer data

You decide what may be tested, when, and when to stop.

Illustrative

A person sets the scope and approves every action; agents do the testing, the watching, and the retesting.

A red-team agent that pentests you on a schedule. It probes your own staging systems and public surfaces with the same techniques the FLASH lists, inside a written scope, and files findings with evidence. Unlike an annual pentest, it runs again when your attack surface changes.

A swarm of watch agents. Each owns one threat: Secrets Watch for keys in public repos, git history, front-end bundles, and images, Login Watch for credential spraying, API Watch for sequential lookups and rate-limit hits, Vendor Watch for new third-party app credentials, Egress Watch for unexpected LLM API calls and outbound data. Small agents are easier to trust, and you add one when a new threat appears. Our cybersecurity agents guide walks through building the first one.

An investigator that connects the dots. When a watch agent flags something, it pulls the related accounts, keys, and logs, and posts one summary in Slack with a recommended action.

A human with the off switch. Blocking an IP, revoking a key, or freezing an account waits for a person's approval. The agent prepares it so approval takes seconds.

Annual pentestAgents on Runtime
How oftenOnce or twice a yearWeekly or after a release, plus continuous secrets checks
CoverageThe apps in the statement of workEvery app in your inventory, including the forgotten ones
Fix verificationNext year's testRetested in the next run
DetectionNot includedWatch agents read logs around the clock
ResponseNot includedRecommended action in Slack, approved by a person

Keep the annual third-party pentest. Your partner bank likely requires it. The agents cover the rest of the year.

How often should agents pentest? Nightly is overkill for most teams: it adds cost, noise, and load on staging without finding much new. Match the cadence to how fast your attack surface changes.

What to testCadence
Secrets in repos, bundles, and imagesContinuously, on every push
Public apps, APIs, and partner portalsWeekly
A change to login, payments, or a new endpointOn that release
Full scope, including legacy and vendor-facing systemsMonthly or quarterly

Which model should run your pentests?

You don't need a frontier model for every run. The FBI says attackers mostly run ARTEX on cheap open-source models like DeepSeek Flash, so testing with the same class of model shows you what they can find. Use a frontier model like Claude or GPT for the hardest findings and the write-up.

ModelGood for
Open-source (Kimi, DeepSeek, GLM)Frequent scans at low cost, seeing what attackers see, and keeping findings inside your own network
Frontier (Claude, GPT)Chaining findings, judging severity, and clear reports for your team

Benchmark before you choose. Run two models against the same scope and the same planted issues, then compare what each found, how many false alarms it raised, and what it cost. Repeat when a new model ships.

Keep it on prem when you need to. Pentest findings are a map of your weak spots. Running an open-source model in your own cloud or data center keeps that map in-house. Runtime works with Anthropic, OpenAI, Google, and open-weight models, and runs in your AWS, GCP, or Azure account or self-hosted.

The stack

JobExample tools
Activity and application logsDatadog logoDatadog
Splunk, Elastic
Cloud and infrastructure logsGoogle Cloud logoGoogle Cloud
AWS CloudTrail, Microsoft Entra audit logs
Code and secrets scanningGitHub logoGitHub
GitLab
The agentsClaude Code logoClaude Code
Codex
Where the team decidesSlack logoSlack
PagerDuty, Jira
Agent infrastructureRuntime logoRuntime

Runtime is the AI agent harness for payment teams. It runs Claude Code and Codex agents on a schedule, on alerts, or when someone tags them in Slack, each in its own sandbox inside your cloud.

Templatesecurity-swarmWeekly pentest, on every push, on Datadog alerts, and when tagged in #security-ops
Skills
  • scoped-pentestTests only the hosts in the signed scope, stops on real customer data, and files each finding with evidence.
  • secrets-watchScans public repos, git history, front-end bundles, and images for keys, and checks whether each one still works.
  • machine-speed-watchFlags bursts, sequential lookups, and failed logins with small changes across your logs.
  • vendor-access-reviewChecks new third-party app credentials, consent grants, and keys that were never rotated.
  • contain-with-approvalPrepares the block or key revocation and waits for a person to approve it in Slack.
Tools
  • Datadog
  • Google Cloud
  • GitHub
  • Slack

Step 1: Write the scope

The scope is the line between a pentest and an attack. Write down which hosts the red-team agent may test, which it may never touch (production payments, third-party systems you do not own), the testing window, and when it must stop and page someone. Have your security lead sign it. This step stays human.

Step 2: Ask for the first recurring pentest

Every Monday at 2 a.m., and after any release that touches login or payments, run a pentest of the hosts in scope.md against our staging environment. Focus on the techniques in the FBI ARTEX FLASH: IDOR, SQL injection, unsigned SAML, missing authentication on APIs, and secrets in front-end code, public repos, and images. Use our open-source model for the scan and Claude for the report. Stop and post in #security-ops if you see real customer data. File each finding with the request, the response, and a severity. Retest last week's findings and mark the fixed ones.

Say why, not just what. "Check every partner portal, because attackers go for the apps nobody watches" lets the agent find the portal you forgot to list.

Name the failure you're preventing. "Never test hosts outside scope.md, even if you find a link to them." One sentence like that keeps a pentest from becoming an incident.

Step 3: Start the watch agents

Start with the one threat you worry about most, usually sequential lookups on login and search. Point the agent at your Datadog and cloud logs with read-only credentials, added through the Runtime dashboard, never pasted in chat. Run it in shadow mode for two weeks and compare its flags with what your team sees.

Step 4: Wire containment to an approval

Let the agents prepare the action, not take it. Every block, key revocation, and account freeze becomes a Slack approval with the evidence attached. Save the logs before anything is revoked.

Step 5: Protect your own agents

Your security agents read attacker-controlled text all day. Treat it as evidence, never instructions. In Runtime, each agent runs in its own sandbox, real keys stay outside it, egress is limited to the hosts you approve, and command deny lists and hooks check every tool call. That also makes the FBI's "unexpected calls to LLM APIs" signal useful: your own agents call only approved endpoints, so anything else stands out. See how Runtime keeps agents contained.

What stays human, permanently

StepWhy
Setting the pentest scopeIt is the legal and ethical line between testing and attacking
Blocking, revoking, and freezingA wrong block stops a real customer from getting paid
Talking to customers, partner banks, and the FBIAccountability cannot be delegated
Deciding what risk to acceptIt is a business decision, not a finding

Lessons that cost the most

The forgotten app is the breach. Inventory first. The agents can only protect what is on the list.

Rotate before you need to. Unrotated keys showed up in the FLASH as the route to forged admin tokens.

Vendors are part of your perimeter. Review third-party app access as often as your own.

Retest everything. A finding that is marked fixed and never retested tends to come back.

Keep the record. When a partner bank asks what you knew and when, the run history is the answer.

A realistic timeline

MilestoneTime
Sandboxed agent environment with scoped credentialsMinutes on Runtime; weeks if you build it yourself
Signed scope, secrets scan, and first weekly pentest on stagingA few days
First watch agent in shadow modeTwo weeks
Containment approvals in Slack, more watch agentsOne week each

The point

ARTEX did not invent new attacks. It made old ones cheap, fast, and parallel. The fintechs that hold up will be the ones that test themselves as often as attackers do, and answer in minutes instead of days. Start with one workflow: a weekly pentest of your own staging environment.

Back to defending.

Sources

Frequently asked questions

What is ARTEX?

ARTEX is an open-source, AI-driven penetration testing framework written in Go. You give it a goal in plain language, and a planner agent splits it into tasks that worker agents run: scanning, exploiting, escalating privileges, and pulling data. It is not an AI model but an agent harness that plugs into commercial or open-source LLMs. It was built for authorized security testing, and attackers have used it in real intrusions since at least August 2026.

Who created ARTEX?

A developer known as Autumn-27 released ARTEX on GitHub in July 2026. The Wall Street Journal, as reported by The Korea Herald, identified the developer as Li Puhua, a Chinese cybersecurity engineer. The developer said it was meant for authorized testing within a defined scope.

Is ARTEX still available on GitHub?

The original repository was made unavailable on October 8, 2026, after reports of abuse. The FBI says ARTEX still spreads through community forks, archived copies, localized versions, and a commercial edition called ARTEX PRO.

Is ARTEX malware?

No. ARTEX is a dual-use security tool, like other pentesting frameworks. It has no CVE and there is no patch for it, because it is not a vulnerability. The risk is that it makes real attacks faster and cheaper, so defenses have to be hardened and faster too.

What did the FBI warn about ARTEX?

FBI FLASH-20261113-001, published November 13, 2026, says criminal and state-aligned actors have used ARTEX since at least August 2026 against financial services and other sectors. Some intrusions went from first targeting to privileged access in under three hours. It cites attacks on South Korean financial institutions that exposed personal data of about 67,000 people, and more than 1,600 exposed ARTEX consoles.

Did ARTEX hit Chase or Visa?

No evidence says so. Chase attributed its November 16, 2026 outage to a software bug in on-premises hardware and said there is no evidence it was cyber-related. Separately, the n0n extortion group listed JPMorgan Chase and Visa as victims; Visa said the incident involved a third-party provider in Asia Pacific and that its own systems were not involved. Neither event has been linked to ARTEX.

How should a fintech defend against AI-driven attacks like ARTEX?

Find your weak spots before attackers do and shorten the time to respond. Inventory every internet-facing app, including partner portals and mobile backends, rate-limit business APIs, rotate long-lived keys, audit third-party app access, and alert on machine-speed patterns like sequential lookups. Scan public repos, git history, and front-end code for leaked secrets. Then run recurring, scoped pentests and have agents read your logs around the clock, with a person approving every block.

How often should a fintech run AI pentests?

Nightly is overkill for most teams. Scan for leaked secrets on every push, test public apps, APIs, and partner portals weekly, test any change to login or payments on release, and run the full scope monthly or quarterly. Keep the annual third-party pentest your partner bank requires.

Can you run AI pentests with open-source models?

Yes. Attackers mostly run ARTEX on cheap open-source models, so testing with models like DeepSeek, Kimi, or GLM shows what they can find, at low cost, and can run inside your own network so findings never leave it. Use a frontier model like Claude or GPT for the hardest findings and the report, and benchmark models on the same scope before choosing.

How do attackers use leaked secrets from public repos?

ARTEX agents search public code for API keys, database URLs, and admin passwords, including in git history, front-end JavaScript, and container images, then log in with them. Microsoft, as cited by GitGuardian, saw one agent go from a GitHub search to reading Azure Key Vault data in five minutes. Revoke and rotate any leaked key; deleting it from the code is not enough.

Can AI agents run recurring pentests safely?

Yes, if they test only systems you own, inside a written scope, from an isolated sandbox with scoped credentials and an egress allowlist. Run them against staging first, log every step, and have a person approve anything that touches production. The same agents can retest each fix so findings do not quietly reopen.

How does Runtime help fintechs defend against agentic attacks?

Runtime is an AI agent harness for payment teams. It runs Claude Code and Codex security agents in isolated sandboxes inside your cloud, so teams can schedule recurring pentests of their own systems with open-source or frontier models, run a swarm of watch agents over Datadog and cloud logs, and route every block or key revocation to a human approval in Slack, with a full record of each run.

Start with one workflow: a weekly pentest

Bring your scope. We'll set up a recurring pentest and your first watch agent in Runtime, with every block approved by your team.