Insights

AI Security

The FBI's ARTEX FLASH asks how long you take to go from detection to containment

The FBI's FLASH on the ARTEX agentic pentest framework says some intrusions went from initial targeting to privileged access in under three hours, and it asks victims to report their time from detection to containment. That is the number to measure in a tabletop before an incident.

November 17, 2026 · 7 min read

The FBI's ARTEX FLASH asks how long you take to go from detection to containment

On November 13, 2026, the FBI released FLASH-20261113-001, which describes ARTEX as an open-source, LLM-driven agentic platform that threat actors have weaponized to automate intrusions at machine speed, from reconnaissance through data exfiltration. The same advisory says some intrusions went from initial targeting to privileged access in under three hours.

The line I keep coming back to is further down. When the FBI lists what it wants in a report about an ARTEX incident, it asks for the time from initial suspicious activity to detection, and the time from detection to containment. Those two numbers decide whether an agentic intrusion ends as an alert or as a breach, and most teams have never measured either one outside a real incident.

A federal FLASH does not make the tool new. It moves machine-speed intrusion from vendor research into the FBI's own guidance, with recommendations a defender can act on this week. I want to walk through those recommendations and map them to three things: test the incident response plan, keep detections on behavior, and be able to respond without waiting on a person.

What the FBI says ARTEX is

The FLASH describes ARTEX as an autonomous, multi-agent penetration testing platform written in Go that turns goals written in natural language into machine commands, with minimal human interaction after the objective is set. It does not ship its own model. The FBI says it integrates with external LLM providers, both frontier models and open-source or Chinese-origin models, and that analysis of attacker-controlled instances found DeepSeek Flash family models most prevalent because of their capabilities, relatively low cost, and lack of observable guardrails when requests are framed as penetration testing.

That is a statement about which models attackers run, not about who the attackers are. The FLASH attributes the campaigns to "a variety of malicious cyber actors" and does not tie ARTEX use to one country or group, so I won't either.

Taking the original code down did not make it go away. According to the FBI, the developer made the GitHub repository unavailable on October 8, 2026, after reports of abuse, and the framework still spreads through community forks, localized variations, archived copies, and a commercial version. The FLASH adds that researchers identified over 1,600 publicly exposed ARTEX console instances as of early November 2026, and that total deployments are likely significantly higher.

Where the FBI says it has been used

Since at least August 2026, the FBI says, ARTEX campaigns have targeted organizations globally in financial services, government, technology, defense, semiconductor, energy, healthcare, manufacturing, telecommunications, and legal. Two campaigns get named. The FLASH cites financially motivated attacks on South Korean financial institutions that exfiltrated personal information on approximately 67,000 individuals, and a supply chain compromise affecting ConnectWise Cloud Backup customers that led to attempted sign-ins at more than 2,250 downstream tenants.

The pattern I would plan around is in the capabilities list. The FBI says ARTEX is good at finding systems that get less security attention than flagship applications: partner and supplier portals, mobile backends, employee applications, sales-support systems, legacy services, and third-party integrations. It also notes that the framework lowers the expertise needed to run complex intrusions, so small groups can run concurrent campaigns at a scale that used to be out of reach.

The FBI is just as direct about why this is hard. The FLASH says ARTEX generates high volumes of forensically noisy traffic that may trigger alerts, but the speed of the attack leaves limited time for a coordinated defender response before the objective is complete. Noisy is not the same as stopped. If an alert fires at minute two and the first human decision happens at minute forty, the alert did its job and the intrusion still finished.

Turn the FBI's recommendations into a tabletop

The recommendations section is grouped into reducing external attack surface, hardening identity and access, strengthening application protections, and preserving evidence while responding. Most of it reads like a checklist. I would run the last group as an exercise instead.

Three of the FBI's response steps pull against each other in a useful way. The FLASH says to preserve sign-in, audit, and SIEM records, but not to delay urgent containment while waiting for exports. It says an appropriately authorized responder should remove unauthorized credentials, grants, and identities after preserving evidence. And it says to keep affected application connections disabled and not to delete enterprise applications as a first response.

That gives you a ready-made inject. A third-party integration in your tenant starts adding credentials and picking up privileged roles. Who is authorized to disable it? How long does it take to preserve the logs? Does someone delete the app in a hurry and take the evidence with it? Run it with a timer and write down the two numbers the FBI asks for. If you have never measured detection to containment, the first exercise gives you a baseline.

The inventory recommendation belongs in the same session. The FBI recommends inventorying every internet-facing application, including partner and supplier portals, mobile backends, legacy services, third-party integrations, and applications created across different business units. A tabletop is a good place to find out that nobody in the room owns half of that list.

Keep detections on behavior

Look at what the FBI asks organizations to monitor: credential additions, privilege assignments, and anomalous app-only activity on highly privileged applications, plus successful and failed service principal sign-ins reviewed separately, with attention to sources outside expected vendor infrastructure rather than only published IP addresses. None of that depends on recognizing ARTEX by name.

The attack characteristics the FBI wants reported point the same way: compressed bursts of activity, high volumes of failed attempts with small iterative changes, sub-agents overwriting each other or generating duplicate requests, and sequential lookups or unusually high numbers of records accessed per session on partner and employee portals. The FLASH also lists outbound connections to LLM API endpoints and unfamiliar LLM proxies or resellers in egress logs as things worth investigating alongside other suspicious activity.

Indicators like IP addresses age quickly, and the FBI itself cautions that ephemeral indicators such as filenames and IP addresses may not indicate a compromise. Behavior is what carries over to the next fork.

Respond at the speed of the attack

The FLASH does not use the phrase autonomous response, and it doesn't need to. If some intrusions reach privileged access in under three hours and your response starts with someone tracking down the right admin, the math does not work.

What I would build is narrow. Pick a few high-confidence behaviors from the list above, such as a third-party application adding a credential and gaining a privileged role in the same window, and decide in advance that the first response is automatic and reversible: disable the connection, revoke its sessions, snapshot the logs. That matches the FBI's guidance to disable rather than delete and to preserve evidence without delaying containment. A person reviews afterward and decides what becomes permanent.

If you report an ARTEX incident to IC3, the FLASH asks for your time from detection to containment. Measure it in a tabletop first, so a real incident is not the first time you learn that number.