← 全部新聞

文章以英文釋出。

Threat IntelligenceAttack SurfaceAI Security

The FBI Says Attackers Are Using an AI Pentest Framework to Go From Target to Privileged Access in Under Three Hours, and It Favours Forgotten Portals. Do You Have a Current List of What You Expose?

Obiguard Research Team·2026年11月17日·5 min read

The FBI has issued a FLASH advisory saying multiple threat actors have used ARTEX, an open-source AI penetration-testing framework, to automate intrusions since at least August 2026, with some going from initial targeting to privileged access in under three hours. According to BleepingComputer's report of 15 November, the 13-page advisory was released on Friday 13 November and includes indicators of compromise. We have not read the advisory itself; this article relies on that report.

We covered the first ARTEX intrusions on 3 November, which focused on how fast the agents move and how loud they are. This article looks at a different line in the FBI's warning: where the framework works best.

What the FBI says

  • It is a framework, not a model. ARTEX is written in Go and has no model of its own. It connects to external LLMs, reportedly including Anthropic Claude, OpenAI GPT, DeepSeek, GLM, Grok and Kimi. Attackers give it goals, and it turns them into commands for specialised agents handling reconnaissance, exploitation, privilege escalation and data theft.
  • One operator, many targets. Because the agents do the routine work, a single operator can run several intrusions at once.
  • It favours overlooked systems. The FBI says ARTEX is especially effective against partner portals, employee applications, legacy services and third-party integrations.
  • It is still available. The original developer reportedly removed the GitHub repository on 8 October after reports of abuse, but it persists through community forks, archived copies and localised versions. The FBI also describes a commercial edition. BleepingComputer cites more than 1,600 publicly exposed ARTEX consoles as of early November and says the real number is believed to be significantly higher.

Campaigns the FBI points to

According to the same report:

  • ConnectWise Cloud Backup. Microsoft attributed the October campaign to Storm-0940. Huntress reported at least 162 compromised tenants, and the attackers tried to authenticate to more than 2,250 downstream Microsoft 365 tenants. We wrote about it in our October article.
  • South Korean banks. A Chinese-speaking actor combined ARTEX with Claude Code agents against Shinhan Bank, KB Kookmin Bank and Hana Bank, exposing personal information of about 67,000 people.
  • South Korean arrests. Police arrested seven suspects linked to attacks on more than 90 organisations, with ARTEX reportedly used in most of them and about 8.2 million records stolen.
  • Vapor Panda. CrowdStrike links this China-based group to attacks on more than 200 organisations in 14 countries, as covered on 3 November.

Why the target list matters

Partner portals, employee tools and legacy services share a trait: nobody owns them closely. They were built for one project, handed to a supplier, or left running after the team moved on. A human attacker has to choose which of those to probe. An agent can probe all of them, in parallel, and at the cost of some tokens.

That changes the defender's arithmetic. Obscurity used to give forgotten systems some protection, because the attacker's time was limited. If the attacker's time is cheap, the forgotten system is simply the one with the least monitoring and the oldest software.

What the FBI recommends

The report paraphrases the advisory's mitigations. The first three are about exposure; the rest are about detection:

  1. Secure internet-facing applications. The FBI's guidance, as relayed, starts with knowing which ones you have.
  2. Restrict access to cloud administration interfaces to trusted networks.
  3. Review application permissions and authentication logs.
  4. Rate-limit sensitive APIs.
  5. Monitor for bursts of automated requests, unexpected credential or privilege changes, suspicious cloud activity and large data transfers.
  6. Treat IP indicators with care. They may be outdated or cause false positives. The report also mentions a third-party detection pack from MalwareMustDie, which is not part of the FBI advisory.
  7. Report suspected ARTEX activity to the FBI or IC3.

The question to take away

If an automated agent scanned your domain tonight, which partner portal, staff tool or old service would it find that your own team has no record of?

How Obiguard Scout fits

The FBI's first step assumes you have an inventory, and most organisations do not have a current one. Obiguard Scout is our external attack surface management product. From a domain you verify, it maps the hosts and services reachable from the internet and what each appears to be running, and flags what changed between scans. That gives you the same view an automated attacker starts with: the forgotten portal, the supplier-run login page, the staging host nobody retired. You can then decide which to patch, restrict or shut down before an agent decides for you.

Scout only sees what is visible from outside. It does not inspect logs, cannot tell you whether a system has already been compromised, and does not review application permissions inside Microsoft 365 or other cloud tenants. The detection steps in the FBI's list still need telemetry from inside your environment. To see what your domain exposes, run a free Scout check on a domain you own, or get in touch.

Sources

Obiguard 如何提供協助

將其轉化為強制執行的政策,而不只是提高意識。

Obiguard 位於貴組織每一次 AI 請求之前——依據護欄、合規框架和稽核軌跡篩查提示詞與輸出,而這樣的報導正說明了它們的必要性。

了解運作方式 →