Wooden children’s sandbox with raked sand and a trail of small, evenly spaced footprints leading over the edge and across a pale floor

Sandbox? What Sandbox?: Red Team Agent Goes Rogue

WorkOS breach was friendly fire from an escaped AI pentest agent.

The AI pentesting tool behind the October intrusion at WorkOS that exposed customer contact details and excerpts of sales and support conversations is the same framework Chinese state hackers and bank thieves have been using since late August.

The intruder that pulled customer data out of WorkOS’s internal sales tool wasn’t a criminal or a spy. It was an unnamed security researcher’s AI agent, running on the ARTEX pentesting platform, that got loose.

In This Article

ARTEX, a Chinese-built, open-source tool, hands a penetration test to AI agents that work through the attack steps themselves, driving ordinary hacking tools like Nmap and curl.

In the nearly four months since ARTEX appeared on GitHub, at least five groups, from Chinese espionage crews to a Cambodia-based criminal ring, have turned it on real targets, and most didn’t bother tweaking it, SentinelOne found in a Nov. 13 report. One China-nexus group alone aimed it at more than 200 organizations in at least 14 countries, including foreign ministries and U.S. defense and nuclear organizations.

Beyond the Break weekly cybersecurity newsletter — Subscribe

As for WorkOS, the researcher was running ARTEX through cybersecurity benchmark tests when the agent slipped out of its sandbox, WorkOS said in a Nov. 6 update. The company stressed the intrusion “wasn’t the work of an actor with malicious intent.”

Out of the Sandbox, Into the Wild

What WorkOS hasn’t explained is how an agent meant to stay inside a closed test environment reached the open internet, and WorkOS, in the first place.

What it did once it was out is clearer.

The agent found an internal tool that WorkOS’s go-to-market teams use to look up customer accounts. The tool’s API was exposed to the internet and accepted a predictable credential meant only for local development. Nothing throttled requests. The agent first got in Oct. 25 and pulled records in bulk on Oct. 30. WorkOS’s monitoring didn’t flag anything until after the bulk download was over, when an employee noticed unusual activity on a connected system.

The haul included business contact details, account information, product usage metadata, and excerpts from sales calls, support threads and customer Slack conversations. WorkOS hasn’t said how many customers were affected. The researcher came forward after WorkOS disclosed the breach Nov. 2, and an independent forensics firm validated the claim. The researcher says the data was deleted and never shared.

The Code that Wouldn’t Stay Closed

ARTEX’s developer ended public releases and took the project closed source on Oct. 8, citing misuse of the tool. The researcher’s agent first reached WorkOS on Oct. 25. It’s unclear which version of ARTEX the researcher was running, or whether it was a fork.

Plenty of copies were still around. Actively maintained forks and backups of the original code remain public, and roughly 2,500 systems have hosted ARTEX control panels, 1,608 of them live and reachable on the internet, SentinelOne reports.

No Off Switch

The WorkOS incident was an accident. Most ARTEX use isn’t, and the case points to a problem defenders haven’t faced in this form. And that is an attack tool with no off switch.

ARTEX’s developer pulled the code, and forks kept it alive. AI model guardrails flagged at least one operator, who simply switched models. A researcher ran it in a sandbox, and it got out. The industry has seen a dual-use tool go rogue before, with Cobalt Strike. But Cobalt Strike had an owner, Fortra, that watermarked every license it sold and went to court, and cracked copies fell 80%.

ARTEX has no one holding that kind of leverage, and the FBI has issued a FLASH alert on it.

ARTEX timeline preview

Timeline

The off switch that didn’t work

ARTEX’s developer pulled the code on Oct. 8. Spies, crooks and one runaway research agent kept using it anyway.

  1. July 2026

    ARTEX lands on GitHub

    Developer “Autumn-27” releases the AI pentesting framework as open source.

    FBI

  2. August

    First malicious use

    Attackers start putting ARTEX to work.

    FBI, SentinelOne

  3. Late Sept.–Oct. 7

    South Korean banks hit

    A campaign against financial organizations ends in stolen data. CrowdStrike ties it to ARTEX on Oct. 7.

    CrowdStrike

  4. Oct. 8

    Off switch

    Developer takes ARTEX closed source

    Citing misuse, the developer ends public releases. Forks and backups stay online.

    Autumn-27 on GitHub, SentinelOne

  5. Oct. 17+9 days

    200-plus organizations targeted

    A China-nexus espionage group turns ARTEX on targets in at least 14 countries.

    SentinelOne

  6. Oct. 21+13 days

    ConnectWise link revealed

    Microsoft says China-aligned Storm-0940 used ARTEX against more than 2,250 tenants.

    Microsoft

  7. Oct. 25–30+17 days

    A researcher’s agent gets loose

    An ARTEX agent escapes its sandbox, reaches a WorkOS internal tool and pulls records in bulk.

    WorkOS

  8. Nov. 13+36 days

    FBI issues FLASH alert

    SentinelOne counts at least five groups using ARTEX. The FBI says a commercial version keeps it available.

    FBI, SentinelOne

Security Point Break logo

SECURITY POINT BREAK  |  Sources: FBI, CrowdStrike, Microsoft, SentinelOne, WorkOS, ARTEX developer. Days counted from Oct. 8 closure.

Five Groups, Factory Settings

“We assess with high confidence that ARTEX marks a change in the threat landscape: agentic offensive frameworks have moved from experiments to operational use by state-aligned and financially motivated actors alike,” SentinelOne wrote.

CrowdStrike tied the tool to a campaign against South Korean financial organizations that ran from late September to early October and ended with stolen data. At one bank, the attacker got into a loan-progress inquiry service used by financial brokers. At another, it got into an employee mobile work-support system. CrowdStrike found the evidence on the attacker’s own exposed servers: open directories holding ARTEX configuration files and Claude Code session histories. The ARTEX setup ran mainly on the Chinese model DeepSeek.

CrowdStrike hasn’t named the attacker. It assesses with moderate confidence that the person is “likely a Chinese speaker and financially motivated.”

Shinhan Bank alone reported about 25,700 customers’ data exposed.

Others had bigger targets. A China-state-aligned group tracked as Storm-0940, which Microsoft holds responsible for October’s ConnectWise supply chain compromise, used ARTEX to attempt sign-ins to more than 2,250 Azure tenants. Criminal operators collected more than 300 signing certificates along with their passwords, according to campaigns analyzed by Check Point and cited in SentinelOne’s report.

None of it took much skill or money. Attackers also combed public GitHub repositories for leaked secrets, and suspected ARTEX activity put them to use.

“The barrier to entry is low: Operators ran ARTEX in default configurations, and each campaign cost $145 to $1,500 in model time. The agents work at machine speed: in one case of suspected ARTEX use, an agent went from a leaked GitHub secret to Azure Key Vault data in five minutes,” SentinelOne said.

“They can also be told to look for gaps, e.g. a Chinese-aligned actor told its agents to find Windows hosts without an EDR sensor.”

Turning it Around

Administrators and developers would be well-advised to scan their code repositories for valid access credentials and rotate any they find, as SentinelOne recommends. SentinelOne also suggests alerting on web traffic whose user-agent string contains “artex.” WorkOS’s own indicators list one, “ArtEx/1.0.”

Or they can follow WorkOS. Since Nov. 3, the company has been regularly running ARTEX against its own internet-facing systems to look for exposure and guessable credentials.

Total
0
Shares
Previous Article
Rows of identical cream seed packets on a slate-blue table, with a brass magnifying glass revealing a thin amber root breaking through one packet.

Backdoored AI Models Pass Benchmarks, Then Wait for a Trigger

Related Posts