Skip to content

The Robots Didn't Invent a New Attack. They Found the One You Forgot.

The FBI's ARTEX alert shows AI agents exploiting old, known weaknesses in hours. Five steps to take now, and how to see what the agent sees first.

Evan Goldberg·VP of Marketing
Published: November 19, 2026·5 min read

Three hours versus 68. That's the gap that matters in the FBI's November 13 FLASH alert on ARTEX, an open-source AI "penetration testing" framework now used by criminal and nation-state crews. The FBI cites intrusions that reached privileged access in under three hours. When Akamai reviewed this autumn's South Korean bank breaches, it found detection at three major banks took anywhere from 15 to nearly 68 hours. By the time most teams notice, the agent is done.

What ARTEX is, the tl;dr

ARTEX is a team of AI agents. An operator types a goal in plain English, and the agents scan, break in, escalate privileges and stage data for theft, often running on cheap models like DeepSeek. Confirmed victims include South Korean financial firms, where data on about 67,000 people was stolen. In another campaign, which Microsoft attributes to Storm-0940, attackers used trusted ConnectWise Cloud Backup apps to target more than 2,250 Microsoft 365 tenants. The original repo came down in October, but the forks and a paid "ARTEX PRO" version are still circulating. ARTEX was built as open source, and that cuts both ways. Anyone can read the code, run it against their own systems and improve it. Attackers get the same access, and a takedown doesn't remove code that has already been forked.

Nothing here is new. That's the problem.

Read the FBI's technique table looking for a futuristic AI exploit. You won't find one. It lists unsigned SAML assertions, SQL injection, admin passwords hardcoded in JavaScript, unrotated signing keys and over-permissioned OAuth apps. Google's Threat Intelligence Group put it plainly: attackers still rely on compromised credentials and unpatched systems. What changed is speed and scale.

Every organization carries a backlog of small, known, unfixed mistakes. Until now, that backlog was protected mostly by the attacker's limited time and patience. In one year, cheap models, multi-step reasoning and point-and-click tooling removed that protection. An agent doesn't get bored on the 500th directory request. It doesn't skip the forgotten loan-broker portal because the main site looked hardened, and that portal is exactly the kind of place where the Korean breaches happened.

If your security strategy depends on attackers not finding something, you no longer have a strategy.

The twist: these attackers are loud

ARTEX isn't stealthy. The FBI describes "forensically noisy" traffic: floods of failed attempts with small changes between them, and sub-agents duplicating each other's requests. In the ConnectWise campaign, one password-spray phase logged more than 151,000 failed sign-ins, peaking near 2,900 a minute. Check Point even caught agents reporting cash-outs that never happened.

The alerts fire. The problem is that nobody can act on them in three hours, especially when the targeted asset isn't on anyone's list.

Now scale that up. According to Picus Security, one China-nexus crew, tracked as VAPOR PANDA, ran ARTEX from a single controller against more than 200 organizations in 14 countries. That's an autonomous penetration test, aimed at you, running all the time, for the cost of model API calls. If you still test yourself once a year, your pen test report isn't a defense. It's a historical document.

Five things to do right now

  1. Find the forgotten front doors. List every domain, subdomain and internet-facing app, including partner portals, mobile backends and whatever another business unit launched. For each one, note the IP addresses, certificates, certificate authorities, autonomous systems, DNS providers and technologies behind it. Compare it with what's actually exposed, not with what the CMDB says.
  2. Hunt your OAuth graveyard. In Entra, search vendor apps by application ID, not display name, and include integrations you've stopped using. Review successful and failed service-principal sign-ins separately.
  3. Watch for servers talking to LLMs. Outbound calls to api.deepseek.com or open.bigmodel.cn from a web server or build box deserve a hard look.
  4. Rate-limit the boring portals. Alert on sequential record lookups and unusually high record counts per session on partner and employee portals.
  5. Lock the admin back doors. Disable basic publishing auth on Kudu/SCM, restrict management interfaces to trusted networks, and rotate any secrets found in front-end bundles or pipelines.

See what the agent sees, first

One of the first things ARTEX does is map what you have on the internet, using WHOIS lookups and scanning. Red Sift Hardenize does the discovery half of that for you, every day, automatically. Starting from a single seed domain, it finds subdomains through certificate transparency logs and passive DNS, then uses the metadata in those DNS records to run reverse lookups across passive DNS that surface more of your domains and subdomains. It assesses the networks and web pages it finds, pulls out any new hostnames, and starts the cycle again. It also connects to your cloud providers, registrars, DNS providers and certificate authorities, so assets nobody remembered launching end up in the inventory too.

For every host, Hardenize checks the DNS, web and email configuration, scans the top 1,000 TCP and top 1,000 UDP ports on each IP address, and records which protocols are exposed and which autonomous systems everything runs on. It tells you which web technologies, certificate authorities, DNS providers, registrars and registries you depend on, and alerts you when your DNS changes.

Then it monitors for the configuration weaknesses that make an easy target, from weak TLS to dangling DNS records ripe for subdomain takeover. Add Red Sift OnDMARC and you see every source sending email as your domains, legitimate or malicious.

Add Red Sift Brand Trust and you see the lookalike domains registered against the ones Hardenize found. It won't replace testing your apps for flaws like SQL injection, but it gives you the external map an attacker's agent starts from, before the agent gets there.

Agentic attacks will only get faster. The organizations that come through this well won't be the ones with the most AI. The industry's reflex will be "fight AI with AI," but even Akamai's researchers put it bluntly: the solution is not simply more AI. The winners will be the ones that know what they own.

Evan Goldberg
Evan Goldberg
VP of Marketing

Evan Goldberg is Vice President of Marketing at Red Sift, bringing more than 15 years of experience building and leading cybersecurity and SaaS marketing teams.