On November 13, 2026, the FBI published a FLASH alert on the ARTEX framework, numbered FLASH-20261113-001. The alert describes ARTEX in these words: “ARTEX is an open-source, large language model (LLM)-driven agentic platform that threat actors have weaponized to automate multiple phases of cyber intrusions at machine speed, from reconnaissance through data exfiltration.” According to the FBI, a variety of malicious actors have used it since at least August 2026, and some intrusions progressed from initial targeting to privileged access in under three hours. The same alert records that the developer stated the tool was intended for authorized security testing within defined scopes and made the original repository unavailable on October 8, 2026, while forks and archived copies keep circulating. U.S. searches for ARTEX surged in the days after the warning.
That leaves a concrete question for the people who already run tools of this kind, from system administrators and managed-service providers to security consultants and students: when does running an AI penetration-testing framework become a federal crime? Under the federal Computer Fraud and Abuse Act, the answer is that the software is not the offense. The line is permission from whoever controls the target system, the scope and end date of that permission, and what the agent actually does once it is inside. The discussion that follows is general in scope: it describes checks useful in comparable situations and makes no findings about the situation of the ARTEX developer or of any person who has used the framework. Every rule below is federal; state law is a separate question not covered here.
I. Is running an AI penetration-testing tool a federal crime?
A. The tool is not the offense: authorization from the system’s owner draws the line
The federal statute is the Computer Fraud and Abuse Act, codified at 18 U.S.C. § 1030, in the 2024 edition of the United States Code read for this article. Subsection (a) lists the prohibited conduct in seven paragraphs. The one that matters most for a scanning and exploitation framework is paragraph (a)(2), which reaches anyone who “intentionally accesses a computer without authorization or exceeds authorized access, and thereby obtains” certain information, including “information from any protected computer.” The statute defines a protected computer to include one “used in or affecting interstate or foreign commerce or communication,” which is why ordinary business servers, cloud tenants and web applications fall inside it.
Read paragraph by paragraph, subsection (a) does not make it an offense, by itself, to download, install or keep a penetration-testing program. Each paragraph requires conduct directed at a computer or at a victim: access without authorization or in excess of authorization, the knowing transmission of code that causes damage, trafficking in passwords with intent to defraud, or a threat made with intent to extort. Two features of the statute keep that observation from becoming a safe harbor. Subsection (b) punishes anyone who “conspires to commit or attempts to commit” an offense under subsection (a), so the analysis changes as soon as a tool is obtained as part of a plan to enter systems the user has no right to enter. And the statute contains no general exception for private security work: subsection (f) covers only lawfully authorized investigative, protective and intelligence activity of law enforcement and intelligence agencies, not consultants or researchers. For a private tester, the only protection the text offers is permission.
The leading interpretation of that permission comes from Van Buren v. United States, 593 U.S. 374 (2021), decided by the Supreme Court of the United States on June 3, 2021. A Georgia police sergeant had used his own valid credentials to run a license-plate search in a law enforcement database in exchange for money, in violation of department policy, and was convicted of exceeding authorized access. The Court reversed, holding that the clause covers someone who enters areas of a computer, such as files, folders or databases, that are off-limits to him, not someone who misuses access he already has for an improper purpose. It described liability under both clauses of paragraph (a)(2) as a “gates-up-or-down inquiry”: “one either can or cannot access a computer system, and one either can or cannot access certain areas within the system.” For anyone running an autonomous testing agent, Van Buren moves the question away from motive and toward gates: what matters is whether each system the agent touches, and each area inside it, was open to the person who launched it. The Court also warned that the government’s broader reading would have criminalized “a breathtaking amount of commonplace computer activity,” and in a footnote it expressly left open whether that inquiry turns only on technological limits or “also looks to limits contained in contracts or policies.” That open question is the reason a written scope document carries so much weight in practice.
Van Buren remains a working precedent in the Supreme Court itself. In Mullin v. Al Otro Lado, No. 25-5 (June 25, 2026), an asylum case decided by the Supreme Court of the United States, the majority quoted Van Buren, 593 U.S. 374, for the presumption that when Congress amends legislation it “intends the change to have real and substantial effect.” The 2026 Court relied on Van Buren for that interpretive presumption, the same presumption Van Buren had used to conclude that Congress’s removal of the word “purpose” from the 1984 version of the computer statute cut against punishing improper motives. That observed citation shows the 2021 decision is still being applied by the Court; it is not a full citator review of every later ruling under the computer statute.
The Justice Department has also narrowed its own charging practice. Its policy in Justice Manual 9-48.000, updated in May 2022, states that federal prosecutors will not charge access “without authorization” unless, among other conditions, the defendant “was not authorized to access the protected computer under any circumstances by any person or entity with the authority to grant such authorization.” It adds that a prosecutor “should decline prosecution if available evidence shows the defendant’s conduct consisted of, and the defendant intended, good-faith security research,” defined as testing carried out in a manner designed to avoid harm and used primarily to promote security. The Department’s announcement of that policy cautioned that claiming to do research “is not a free pass for those acting in bad faith.” Two limits matter. The Manual is guidance for federal prosecutors, and it says plainly that its principles “are not intended to, do not, and may not be relied upon to create a right or benefit, substantive or procedural, enforceable at law by a party to litigation with the United States.” It does not speak to a state prosecutor or to a company suing under the civil provision discussed in Part II. A tester who relies on good faith instead of documented permission is relying on discretion, not on a defense.
B. Scope, end dates and third-party systems: where authorized testing becomes unauthorized access
Permission has a source, a scope and an end, and a recent appellate decision shows how much turns on each. In United States v. Eddings, No. 23-3017 (3d Cir. Dec. 9, 2025), a precedential opinion of the United States Court of Appeals for the Third Circuit, a worker hired to help organize a fundraiser had been given a link that let her use a board member’s email account. After she declined to continue the work, she still had the link, read the account, downloaded internal documents and sent them to a friend, who was convicted as her accomplice and co-conspirator. The government’s theory was that the worker’s resignation ended her authorization. The court rejected it, vacated the conviction and remanded with instructions to enter a judgment of acquittal. “Whether an employee is authorized to access her employer’s computer is up to her employer,” the majority wrote, and “absent an applicable contract or policy, the employer must act” to withdraw that permission, although revocation “might take nothing more than an email notifying the employee she is not entitled to access the system anymore.” The court also separated the two concepts that autonomous tools tend to blur: “Authorization is permission to access a computer, not access alone.” Applied to a penetration test, Eddings means the authorization letter should come from the party that actually controls each target, state when permission ends, and be honored the moment the client withdraws it in writing, because a working credential or an open port proves access, not authorization.
Eddings is a precedential decision binding within the Third Circuit, and it drew a dissent from Judge Montgomery-Reeves, who would have let a jury weigh a resignation together with the other facts. The majority stressed that its holding is narrow: it did not decide whether firing an employee revokes authorization, and it did not rule out that an employer may provide by contract that resignation ends authorization. For a testing engagement, that narrowness is the practical lesson. A statement of work that says testing authority ends on a fixed date, or when the client sends written notice, gives both sides a clear rule. An open-ended engagement leaves the question to argument after the fact.
Scope is where autonomous agents create the most exposure. The FBI alert reports that ARTEX activity included cases where the agent “abused cross-account cloud trust relationships from an external partner tenant,” pivoted through compromised third-party vendor credentials and trusted business-to-business federation, and that agents can systematically map trust relationships between organizations. Under the reasoning of Eddings, only the party that controls the right of access can grant or withdraw permission. A client can authorize testing of its own web applications and tenants; it cannot supply the permission of a supplier, a software-as-a-service provider or a partner whose systems sit on the other side of a federation trust. An agent that follows a token or a trust relationship into such a system may be on a computer whose owner never agreed, even if the client was eager for the test.
The practical checks follow from that structure. The authorization document should list the assets by host name, address range, cloud account or tenant identifier, and it should say which actions are allowed: scanning, exploitation, credential use, file writes, account creation, password resets, or none of these. The agent should be configured with an allow-list that matches the document and a stop condition that halts it when it reaches anything outside the list. Credentials discovered during the test should be used only where the document permits, since a working password does not convert a third party’s system into an authorized target. Because the FBI alert notes that ARTEX relies on outside language-model providers and flags outbound calls to their interfaces as a warning sign, the engagement should also say whether client data may be sent to an external model at all. Finally, the tester should keep complete logs. If anyone later questions the work, the logs and the signed scope are the evidence of what permission existed and what the agent actually touched.
II. What the federal statute punishes, and what a targeted business can do
A. Damage, autonomous agents and the penalties Congress set
Access is only half of the statute. Paragraph (a)(5) adds three damage offenses, and they are the ones most likely to matter when an agent does more than look. Paragraph (a)(5)(A) reaches anyone who “knowingly causes the transmission of a program, information, code, or command, and as a result of such conduct, intentionally causes damage without authorization, to a protected computer.” Paragraph (a)(5)(B) covers intentional unauthorized access that recklessly causes damage, and paragraph (a)(5)(C) covers intentional unauthorized access that “causes damage and loss,” with no separate mental-state requirement stated for the damage itself. The statute defines damage as “any impairment to the integrity or availability of data, a program, a system, or information.” The FBI alert lists behavior observed in ARTEX intrusions that could fall within that definition: test writes and deletions in cloud storage, web shells planted on servers, accounts created with administrator flags and account takeovers through password resets. Whether a given step amounts to damage under the statute is a question for a court on the facts; the definition is broad enough that a tester should treat every write, deletion or configuration change as something the authorization document must expressly allow.
No decision read for this article addresses an autonomous AI agent, and none should be presumed. The statute’s mental-state words, “intentionally” and “knowingly,” attach to the person. In Eddings, the Third Circuit reasoned that until permission is withdrawn, a worker may not know her authorization has ended and therefore may lack the state of mind the statute requires. With an agent, the evidence of what its operator knew and intended will usually be the objective the operator typed, the scope configured in the tool, the targets loaded into it and the logs of what it did. That is one more reason the paperwork and the configuration should match.
Subsection (c) sets the penalties, and they depend on the paragraph and the harm. A first offense under paragraph (a)(2) carries up to one year of imprisonment, but up to five years when the offense was committed for commercial advantage or private financial gain, in furtherance of another criminal or tortious act, or when the value of the information obtained exceeds $5,000. A first offense under paragraph (a)(5)(A) carries up to ten years when it caused one of the harms listed in the statute, which include loss to one or more persons aggregating at least $5,000 in value during any one-year period, a threat to public health or safety, and damage affecting ten or more protected computers during any one-year period. Paragraph (a)(5)(B) carries up to five years with the same listed harms, and other paragraph (a)(5) offenses carry up to one year. Under subsection (i), a court sentencing a person convicted under the statute must also order forfeiture of personal property “used or intended to be used to commit or to facilitate the commission of such violation.” For an operator, that can include the servers and laptops that ran the agent. The FBI alert reports that researchers identified more than 1,600 publicly exposed ARTEX console instances as of early November 2026, a reminder that the infrastructure running such a tool is itself visible to investigators.
One more federal provision belongs in this part. The FBI alert asks organizations to report console artifacts, configuration files and access logs, and anyone whose systems appear in such records may be asked questions by federal agents. Under 18 U.S.C. § 1001, anyone who, in a matter within the jurisdiction of the federal executive branch, knowingly and willfully “makes any materially false, fictitious, or fraudulent statement or representation” faces up to five years of imprisonment, with an eight-year maximum in the categories the section lists. That offense stands apart from the computer statute. An administrator who learns that a server under his control hosted an exposed console has every reason to answer accurately or to decline to answer through counsel, never to guess.
B. For a business facing an AI-driven intrusion: the civil claim and the two-year clock
The computer statute is not only a criminal law. Under subsection (g) of 18 U.S.C. § 1030, a person who suffers damage or loss because of a violation may bring a civil action against the violator for compensatory damages and for injunctive or other equitable relief. The civil action is available only if the conduct involves one of the first five factors listed in subsection (c)(4)(A)(i), the most common being loss aggregating at least $5,000 during a one-year period, and damages for a violation involving only that factor are limited to economic damages. The statute defines loss as “any reasonable cost to any victim, including the cost of responding to an offense, conducting a damage assessment, and restoring the data, program, system, or information to its condition prior to the offense,” along with revenue lost and other consequential damages caused by an interruption of service. Subsection (g) also sets the deadline: “No action may be brought under this subsection unless such action is begun within 2 years of the date of the act complained of or the date of the discovery of the damage.”
Whether response costs count as loss after Van Buren was the question in Moxie Pest Control (Utah), LLC v. Nielsen, No. 24-4076 (10th Cir. Jan. 21, 2026), a published decision of the United States Court of Appeals for the Tenth Circuit. The pest-control company alleged that a competitor’s sales executive used login credentials obtained from one of its workers to copy confidential sales data from its password-protected system, and it claimed more than $5,000 in investigation costs. At the motion-to-dismiss stage, the court treated the alleged use of a “pilfered username and password” as access without authorization, a point the defendants did not dispute. The district court had dismissed the claim on the theory that Van Buren limited loss to technological harm. The Tenth Circuit reversed, holding that “the CFAA does not require plaintiffs to plead loss from a technological harm,” and that costs reasonably necessary to respond to the offense, such as identifying the perpetrator or the method of access, can qualify. It drew a line, however: the costs of investigating how a competitor used the information after obtaining it are business harms, not statutory loss. For an organization responding to an intrusion, Moxie means the forensic invoices for finding the entry point, scoping the compromise and restoring systems should be documented separately from the costs of measuring commercial fallout, because only the first category reliably counts toward the $5,000 threshold in that circuit.
Moxie is the law of the Tenth Circuit, and the court itself acknowledged that other courts, including a 2022 footnote of the Ninth Circuit, have read Van Buren as requiring technological harm. A business outside the Tenth Circuit should therefore assume that the characterization of its costs may be contested and keep the record detailed enough to show actual impairment where it exists. Moxie also illustrates a point that matters to every company whose credentials leak: logging in with a username and password that were never issued to the person using them is the kind of entry the court treated as access without authorization, whatever tool makes the login.
The practical sequence after an intrusion follows from these sources. First, preserve the evidence the FBI asks for: its alert requests authentication and access logs, application and interface activity, records of outbound connections to language-model interfaces, cloud and container activity, and loss and impact figures. Second, report the incident to a local FBI field office or to the Internet Crime Complaint Center, as the alert encourages. Third, calendar the two-year limit in subsection (g) from the date of the act or the date the damage was discovered, whichever applies on the facts, and do not assume a later start date without advice. Fourth, look at access granted to former contractors, testers and employees. Eddings shows that, at least in the Third Circuit, a person’s continued use of an account the organization never shut off may not be “without authorization” absent a contract or an affirmative step; revoking credentials and sending a short written notice when an engagement ends removes that argument. Where personal data was exposed, the questions shift to notification duties and to individual claims, which are covered in our earlier explainers on what courts require before you can sue after a data breach and on who must notify you when a vendor flaw exposes your data.
Conclusion
An autonomous penetration-testing framework does not change the federal rule; it changes the speed at which the rule can be broken. The Computer Fraud and Abuse Act does not punish owning or running a testing program on systems you are entitled to test. It punishes entering systems or areas of systems that are closed to you, causing damage without authorization, trafficking in passwords and using threats, along with attempts and conspiracies to do those things. Van Buren frames the inquiry as gates that are either up or down, Eddings ties authorization to the party that controls the system and to the moment that party withdraws it, and Moxie confirms that a business can recover reasonable response costs when someone walks through a gate that was closed.
For testers, the decisive documents are a signed authorization from the party that controls each target, a written scope that lists assets and permitted actions, an end date, and logs that show the agent stayed inside those limits. The Justice Department’s good-faith research policy is real but discretionary, and it does not govern a civil plaintiff. For organizations, the priorities are to preserve logs, report the intrusion, separate response costs from business losses, revoke access when engagements end, and keep the two-year civil deadline in view. Speed is the new variable in these intrusions; permission remains the line the law draws.