Your Frontend Bundle Is a Credential Store, and Agent Fleets Know It
Your Frontend Bundle Is a Credential Store, and Agent Fleets Know It
On 5 November, Check Point Research published an analysis of ARTEX, an open-source agentic penetration testing framework, and the criminal campaigns running on it. The headline number is attention-grabbing: one operator recovered more than 190,000 sets of game account credentials for an estimated $180 to $360 in model costs. The more useful detail for a business is how the agents got them. Not a clever exploit chain. A plaintext credential file sitting in a web root, and hardcoded administrator passwords inside frontend JavaScript bundles.
That is the shape of the problem now. Agent fleets do not make attackers smarter. They make it cheap to check every boring thing, everywhere, until something is open.
The fleet is the product
ARTEX is a Chinese open-source tool built for authorized penetration testing: a multi-agent system that plans tests, runs security tools, and records evidence, with a web dashboard for the operator. It was built for work where the operator is expected to stay in the loop. The campaigns Check Point analyzed did the opposite.
The researchers counted 2,557 systems that were hosting or had hosted ARTEX panels, with 1,608 live and publicly reachable at the time of the research. Ten of those were misconfigured badly enough to expose the full ARTEX database, which is how the researchers got a rare look at the operator consoles: which models were configured, how the agents were prompted, and exactly what they did.
What the prompt logs show is a job that has become mostly operational. Beyond the initial goal, the agents ran on their own. The operators mostly asked the same question over and over: what is the progress. In one gaming account campaign, the console ran for about five days on DeepSeek V4.1 Flash and processed roughly 116.8 million input tokens on cache misses, 1.907 billion on cache hits, and 30.8 million output tokens. Check Point estimated the cost at $180 to $360 at official list prices.
That is the new unit economics. A campaign that reaches 13 sites, pulls 190,000 credentials, and tests a sample to find roughly a 3.3 percent hit rate, about 6,000 still-valid accounts, costs less than a mid-range laptop.
The same pattern, three times
The gaming campaign is one of three Check Point described. The others show the same mechanics pointed at different targets.
The Apple certificate campaign targeted 10 gray-market iOS signing shops and ran two consoles in parallel for roughly two weeks, from 14 to 29 October. Across 36 billion tokens on DeepSeek V4.1 Flash and Kimi K3.1 Pro, the estimated cost was $700 to $1,500. The agents found more than 300 code-signing certificates and their passwords. After automated validation, three were usable.
The crypto campaign ran for four days across 292 payment domains for an estimated $145 to $300. It produced 230 findings and 400,791 agent messages. It moved no money at all. Some agents reported successful cashouts that turned out to be hallucinations, based on API responses that said "success" while the wallet stayed at zero.
The exploited vulnerability classes, in Check Point's list:
- Hardcoded administrator passwords in single-page app JavaScript bundles
- Missing authentication on web APIs
- Remote code execution through command injection
- Insecure direct object references
- Weak administrative credentials
- Remote code execution through server-side template injection
- A 21.5 MB plaintext credential file left in a public web root
None of these require a zero day. They are the misconfigurations that survive because nobody has the time to sweep every endpoint, every bundle, every forgotten admin path. An agent has that time.
What happens when a state actor rents the same fleet
Check Point's criminal campaigns are opportunistic. Two other reports from the same period show the same tool inside targeted espionage.
CrowdStrike reported in October that an actor it assesses as likely a Chinese speaker and financially motivated used ARTEX alongside large language models against South Korean financial organizations, with activity from late September to early October 2026 and exfiltrated data. CrowdStrike found open directories containing Claude Code session histories, ARTEX configuration files, and Claude memory files, with DeepSeek v4.1-flash as the primary model backend.
Microsoft reported on 21 October that Storm-0940, a China-state aligned actor it tracks, used ARTEX from as early as September 2026 to automate attacks against hundreds of organizations across aerospace, automotive, chemicals, energy, financial services, government, legal, manufacturing, media, and other sectors. During the ConnectWise supply chain compromise, the actor used the framework to accelerate scanning, password spraying, and OAuth application abuse. In several cases the time from initial targeting to privileged access to cloud infrastructure was under three hours. Microsoft also notes that the majority of the targeting was not successful, which is the honest version of the story: high volume, low hit rate, and the hits arrive faster than a human team can respond.
What the agents still cannot do
It is worth keeping the failures in view, because they define where the human work remains.
The crypto campaign generated 230 findings and moved zero funds. Some of its "successes" were model hallucinations dressed up as validated exploitation. The Apple campaign pulled 300 certificates and produced three that actually passed a validation checklist. Across the ARTEX data, the agents were strongest at breadth and persistence: auditing every endpoint, testing every payload, working through a list of targets without stopping. They were weaker at knowing when a win was real.
That is a familiar division of labor in business automation too. Agents are good at exhaustive checking and bad at final judgment. The difference here is that the checking is now aimed at your perimeter, and the volume is the attack.
What to do with this
The defensive response is not another AI product that promises to detect AI attacks. It is the unglamorous work of shrinking what an exhaustive sweep can find, done at the same scale as the sweep.
- Inventory the whole perimeter, not the documented one. Include staging hosts, old admin panels, forgotten subdomains, and any endpoint that answers without a session. The ARTEX operators found targets by spraying a scope, and they did not care whether the target was in anyone's asset register.
- Sweep web roots and bundles for secrets. A frontend bundle is readable by anyone with a browser. Any key, token, or admin credential that ships in it is public. Move secrets to a store the client never sees, and rotate anything that has been sitting in a build.
- Close unauthenticated APIs, or put them behind real auth and rate limits. Then test object references by hand: change an ID in a request and see whose data comes back.
- Alert on breadth, not only on signatures. One host hitting hundreds of endpoints in minutes is a different signal from a known exploit string. If your logging can only match yesterday's attack, the fleet will look like noise until it is not.
- Scope credentials so one win is not a tour. The campaigns that succeeded used a single opening to reach more. Short-lived tokens, least privilege, and separate credentials per service turn a breach into a contained incident.
- Cap and log your own agents the same way. If you run agent fleets, budget limits, action logs, and approval gates for irreversible steps are the difference between an operator and a liability. Check Point found thousands of exposed ARTEX consoles; the operators were as exposed as their targets.
The pattern is not that AI makes every company a target. Most opportunistic campaigns fail. The pattern is that the cost of trying has collapsed, so the number of attempts against your public surface goes up whether or not anyone has decided to target you. The businesses that come through it are the ones whose perimeter is boring to an agent: authenticated by default, empty of secrets, and logged well enough to notice a machine reading it.
If you are mapping where agents, automation, and your public surface meet in your own operation, talk to us.
Sources
- Hacking on Autopilot: Inside ARTEX Cybercrime Campaigns (Check Point Research, 5 November 2026)
- Unknown Threat Actor Uses AI-Driven ARTEX to Target South Korean Finance (CrowdStrike, 7 October 2026)
- Storm-0940 uses ARTEX agentic AI in ConnectWise supply chain compromise (Microsoft Security Blog, 21 October 2026)
