# How ARTEX Works: A Defender's Guide to Agentic Attacks ARTEX coordinates AI agents across applications, cloud services and identities. Learn what defenders can check safely, with a free infrastructure self-test kit. Author: J.A. Watte Published: November 3, 2026 Source: https://jwatte.com/blog/how-artex-agentic-pentest-framework-works-defense/ --- **ARTEX was an open source penetration testing suite developed in China that attackers used to coordinate AI agents across real organizations. Businesses with exposed applications, weak identity settings, old credentials or poorly controlled service providers face the same classes of weakness described in the reports. The biggest lesson is that familiar defects become harder to manage when one operator can pursue many paths at once. This week, inventory your exposed services, confirm who patches them, and have your IT provider review identity controls, endpoint coverage and old keys using the kit below.** [CrowdStrike's architecture and findings](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/), [origin and later closure reporting](https://www.bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs/). ## What ARTEX was, and what we can verify A penetration test is an authorized attempt to find and validate security weaknesses within agreed boundaries. ARTEX automated parts of that work. An attacker can aim the same capabilities at systems without permission, which changes the activity into an intrusion. [CrowdStrike's October 7 report](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/) calls ARTEX a recently released, open source agentic pentesting tool developed in China. [BleepingComputer's reporting](https://www.bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs/) also describes its original public availability and subsequent closure. We could not verify an exact original release date from the public material reviewed. The developer's public GitHub profile contains a [statement dated October 8, 2026](https://github.com/Autumn-27/Autumn-27/blob/main/README.md#L5-L18). In our translation, it says the intended use was authorized security testing and research, condemns unauthorized use, and announces that the project will become closed source with no further updates or public maintenance. The exact moment every copy became unavailable remains unverified. The profile README contains the closure statement without architecture documentation. We did not find a public ARTEX repository in the developer's repository listing during this review. We therefore use the vendor investigations for the architecture below. We did not obtain or run the framework, derivatives or malware. The reports also concern different operations. The unattributed South Korean financial campaign in [CrowdStrike's first report](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/), VAPOR PANDA in its [October 29 report](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/), and Microsoft's [Storm-0940 investigation](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/) should not be merged into one attacker story. CrowdStrike explicitly distinguishes VAPOR PANDA from Storm-0940. [BleepingComputer's October 26 report](https://www.bleepingcomputer.com/news/security/former-pentester-used-artex-ai-to-steal-82-million-data-records/) adds another account: local media, citing South Korean police, described a former pentester and an operation affecting more than 90 organizations, with an estimated 8.2 million stolen records. We have not independently measured those reported police claims. Hospitals, educational organizations and other ordinary businesses appear in that reporting. This is why an owner should pay attention even without operating a bank. ## A coordinator, specialists and a shared notebook Agentic AI means software that can choose and carry out a sequence of tool-assisted actions toward a goal. A chatbot may explain a firewall setting. An agent can be given access to inspect configuration, compare observations and decide what to examine next. Its actual reach depends on the tools and permissions supplied to it. A large language model, or LLM, is the model interpreting instructions and producing responses. The LLM backend is the model service behind the agent, much as different engines can power the same machine. A tool call is a request from the agent to a program that performs an action. MCP, the Model Context Protocol, is one way software exposes tools to assistants. The reviewed evidence does not establish that this ARTEX deployment used MCP. [CrowdStrike describes](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/) a main agent coordinating subordinate planner and worker agents over a shared graph of facts, findings and goals. Think of a supervisor, several specialists and a linked case notebook. A graph records relationships, such as which account belongs to which application, rather than just a pile of disconnected notes. The report calls individual engagements ARTEX “tasks.” Several engagements could proceed simultaneously. A human operator still steered agents at decision points. This is a person directing a collection of software workers, with automated activity between interventions. The diagram summarizes those relationships without specifying internal message routing.
A human operator steers a main agent that draws on model services and directs planner and worker agents over a shared graph of facts, findings and goals against many targets. Below, eight defensive checkpoints for a small business.
The upper panel summarizes CrowdStrike's reported architecture. The lower panel maps defensive controls to the exposed services, identities and credentials the reports describe.
```text Operator with human instructions ▼ Main agent ← LLM backends ▼ Planner agents + worker agents ▼ Shared graph: facts, findings, goals ▼ Many targets in parallel Defensive checkpoints ├─ Remove unnecessary exposed services ├─ Patch appliances; protect remote access ├─ MFA + validated SAML signatures ├─ IMDSv2 + limited cloud permissions ├─ LAPS + reviewed delegation └─ EDR everywhere + rotated keys + alerts ``` In the VAPOR PANDA operation, CrowdStrike identifies DeepSeek plus a second commercial LLM whose identity it could not confirm. It estimates planning agents consumed on the order of **1.4 billion input tokens in one intrusion**. Tokens are pieces of text processed by a model. That figure describes model input, which can include repeated context. It cannot establish unique stolen record counts or successful attack counts. [Campaign analysis](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). The earlier South Korean investigation reports DeepSeek v4.1-flash as the primary ARTEX backend, with GLM-5.3 and Grok 4.6 supplementing separate Claude Code sessions. These are the investigators' reported model labels. They do not establish that every ARTEX installation uses those models or that every session run through the Claude Code client used an Anthropic model. [CrowdStrike's first report](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/). Microsoft separately describes a multi-agent framework written in Go, with a Next.js user interface, accumulating assets and weaknesses into a graph. Defenders can understand the implications without reproducing those software components. [Microsoft's architecture description](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/). ## What changes for your IT team The main change is the amount of work an operator can keep moving. CrowdStrike observed agents pursuing paths across applications, cloud services, containers, identity systems and mail, then trying alternatives when an approach failed. Fixing the first visible weakness does not establish that the investigation is over. [Observed behavior](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). Parallel work also changes staffing economics. The reports support the conclusion that a small team can cover more targets with less manual effort. They do not provide a verified cost per intrusion. “Cheap scale” describes reduced human effort per attempted path. The computing bill remains unknown. Counts need careful reading. CrowdStrike says the VAPOR PANDA activity beginning October 17 targeted more than 200 organizations across at least 14 countries during an 11-day window. It confirmed post-exploitation activity at nine organizations. Keep the targeted organizations and confirmed compromises separate when assessing these figures. [Primary report](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). Microsoft reports several Storm-0940 intrusions reaching privileged cloud access in less than three hours. It also says most targeting was unsuccessful. Fast attempts can generate plenty of logs while leaving little time to interpret them. The practical question is whether someone can receive an alert, verify it and contain access during your closed hours. [Microsoft's observations and limits](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/). ## The weaknesses, translated into business decisions The following control families organize the defensive review. CrowdStrike's [VAPOR PANDA report](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/) supplies the incident examples and core recommendations. The checks are our defensive application of those findings. ### Public applications and appliances An internet-facing application accepts connections from outside your business, such as a customer portal or remote-access gateway. An appliance is a packaged system such as a firewall, mail gateway or VPN device. Firmware is its underlying software. These systems need a named owner and supported updates even when they have operated quietly for years. CrowdStrike recommends patching internet-facing appliances, particularly long-lived Linux servers. Ask for an inventory showing vendor, version, support status, exposure and last patch verification. “The provider handles it” needs a report behind it. Remove unnecessary exposure and retire unsupported equipment through a planned change. A web shell is an unauthorized server-side program that gives an intruder continuing access. Owners should ask for monitoring of unexpected application file changes and unusual server activity. Finding one calls for incident response and investigation of the access behind it. ### Identity assertions and cloud metadata SAML, Security Assertion Markup Language, carries a statement from an identity provider to an application saying who signed in. Think of a visitor badge issued by a trusted reception desk. Signature validation is how the application verifies the badge really came from that desk. CrowdStrike reports an unsigned assertion being accepted. Ask both your identity provider and application vendor to document rejection of unsigned or invalid assertions and validation of the intended recipient, issuer and time limits. Multi-factor authentication, or MFA, adds another proof of identity beyond a password. It remains valuable, but MFA at the login screen cannot repair an application that accepts an invalid identity statement. These are separate controls. XXE, XML external entity processing, is a flaw in how software handles structured XML documents that can expose information or cause unintended requests. The defensive work is reviewing parser settings and supported patches with the vendor. Owners should not improvise malicious documents to test production software. The cloud Instance Metadata Service supplies information and sometimes temporary credentials to a virtual server. AWS IMDSv2 adds a session-token requirement to metadata requests. CrowdStrike recommends requiring it and restricting cloud roles' administration permissions. A role is an assigned set of permissions, like a job description for software. Requiring IMDSv2 helps reduce a known class of exposure; it does not make an excessively powerful role safe. [AWS metadata settings](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/configuring-instance-metadata-options.html). ### Container administration and Windows identities A container packages an application with what it needs to run. Orchestration software manages groups of containers. Its administration interface can have extensive control over production workloads. CrowdStrike reports abuse of an orchestration service without authentication. Review the service's configuration, network exposure and required identity checks. An internal address alone does not establish safety after another machine is compromised. Active Directory is Microsoft's directory of organizational computers, users and permissions. Windows LAPS, Local Administrator Password Solution, manages local administrator passwords so machines need not share one reusable password. CrowdStrike reports attackers reading LAPS passwords. That is a reason to verify who can retrieve them and whether the directory is protected, rather than to abandon LAPS. Delegation lets one service act on a user's behalf when communicating with another service. Resource-based constrained delegation is a version where the receiving resource specifies who may do that. It resembles an office authorizing a courier to represent particular staff. Unnecessary delegation or the ability to change it can expand an intruder's reach. Have your directory administrator review delegation settings, their business purpose and who can modify them. The kit avoids retrieving passwords. ### Old exports, mail and remote access An export or backup can preserve credentials long after a system changes. CrowdStrike describes authentication keys in a 2022 mail-server directory export that likely had not been rotated, plus VPN configuration from sent mail and a client key described as coming from a May 2024 leak. Those dates belong to that reported incident. They illustrate why deleting an exposed file does not revoke a copied credential. [Incident account](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). A VPN provides a protected connection into a network. An RMM tool, remote monitoring and management software, lets an IT provider administer devices. Inventory both, including abandoned installations and vendor integrations. Verify supported versions, MFA, named accounts and restricted administrative access. For suspected key exposure, preserve evidence and have the owner rotate or revoke affected keys, certificates and dependent credentials in a controlled recovery plan. ## The service provider route matters even without a public server A supply chain compromise reaches customers through a supplier or trusted component. Microsoft's [October 21 Storm-0940 report](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/) describes abuse of compromised ConnectWise cloud application identities already trusted by customer Microsoft 365 environments. Application access did not require an interactive user login. OAuth is a system for granting applications permission to use resources. A service principal is an application's identity inside your organization's cloud directory. Microsoft Graph is an interface applications use to work with Microsoft 365 data and directory resources. These software identities deserve review alongside employee accounts. Ask your IT provider for an application-permission inventory, service-principal sign-in review and checks for unexpected credentials, owners or privilege changes. Microsoft warns that later unauthorized changes may survive remediation of the original application. It also cautions that successful authentication alone does not prove data theft. Keep those distinctions in the incident report. The report says ARTEX attempted sign-ins to more than 2,250 tenants, meaning separate customer directory environments. Do not mistake that for a confirmed victim count. Follow current vendor containment and restoration guidance if affected; removing application records before preserving evidence can make investigation harder. [Microsoft's containment guidance](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/). ## What to look for in your own records EDR means endpoint detection and response: software and associated monitoring that record suspicious device activity and support investigation. Compare your device inventory with the security console's recently reporting hosts. Include servers and remote devices. A license count or a list of installed agents cannot establish that every host is reporting useful data. An IOC, indicator of compromise, is an artifact associated with known malicious activity, such as a domain or file hash. It provides a lead that needs investigation and corroboration. CrowdStrike published ARTEX-related markers: `artx` or `artex` in command-line parameters, cloud role names beginning `artx`, administration-command comments beginning `art18-`, and an `artx5v` directory beneath Windows ProgramData. [Marker recommendations](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). The kit includes defanged domains and addresses from both reports. Defanging replaces a dot with `[.]` so readers do not accidentally follow an address. Search existing records; do not visit those destinations. Match dates, host activity, identities and approved work before drawing a conclusion. Names can change, and shared infrastructure can produce unrelated matches. MITRE ATT&CK is a catalog of adversary behavior with identifiers that help different security teams discuss the same technique. CrowdStrike maps this activity to entries including T1190 for public-facing application exploitation, T1552.005 for cloud metadata credentials and T1555 for password stores. Ask your monitoring provider what evidence it collects for those behaviors. A product saying it “covers ATT&CK” is less useful than showing which events it would retain and who investigates them. [Published mapping](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/). CrowdStrike also reports a backdoor named BLOODALCHEMY, meaning malware that provides continuing unauthorized access. Elastic's [October 2023 analysis](https://www.elastic.co/security-labs/threat-command/disclosing-the-bloodalchemy-backdoor) explains that malware family but does not mention ARTEX in the page reviewed. CrowdStrike supplies the later campaign connection; Elastic supplies historical malware context. ## What Anthropic says, and the legitimate use of these tools CrowdStrike found Claude Code histories, ARTEX configuration and Claude memory files in attacker-controlled open directories. An open directory exposes a server's file listing to visitors. That accidental exposure provided investigative evidence and shows why your own agent instructions, histories and credentials belong outside a public website's published files. [South Korean campaign report](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/). Anthropic's [September 2026 threat report](https://www.anthropic.com/threat-intelligence-report-september-2026) describes misuse of Claude by criminals and other actors, account disruption and changes to safeguards. Its [usage policy](https://www.anthropic.com/legal/aup) prohibits compromising systems without authorization while allowing lawful authorized security work. Neither page supplies an ARTEX-specific response to these October campaigns. We did not verify such a statement and do not infer that Anthropic endorsed the activity or confirmed every vendor observation. Our [Cloudflare security audit skill article](/blog/cloudflare-security-audit-skill-small-business/) describes legitimate agents dividing review work and challenging findings. The technological family is similar. The difference is authorization, scope and containment: whose code or systems may be examined, what actions are allowed, where data goes and who reviews results. A contained source review can help find defects without giving an assistant unrestricted production access. A deployed-system penetration test requires its own permission and controls. ## How to test your own infrastructure against this class of attack Begin with [written authorization](/downloads/ai-security-audit-authorization-kit.md). Identify systems you own, systems a provider owns, and shared services whose terms constrain testing. Our public scanning rule is passive only. Anything active requires an engagement with written scope. The downloadable worksheet maps each reported weakness to an inventory item and a read-only check. Ask for evidence: the cloud metadata setting, a vendor's SAML validation statement, the LAPS event status, an approved delegation list, a device coverage comparison. Record “unknown” when the person doing the review lacks permission or the required records. A completed spreadsheet still leaves untested behavior and unknown attack paths. Use findings to assign changes with owners and retest dates. Public exposure without a business need, an unsupported remote-access appliance or a suspected exposed signing key deserves prompt attention. Other configuration changes need staging, compatibility review and a rollback plan. The kit's inspection commands make no changes; remediation is separately approved work. For deeper validation, ask your existing IT provider for a specialist referral or approach an independent penetration testing firm with relevant cloud and identity experience. Request an anonymized sample report and references for comparable environments. Scope should specify assets, dates, permitted techniques, exclusions, data handling, stop conditions, emergency contacts and whether AI services may receive any customer material. Explicitly address suppliers and shared infrastructure. Require a report with an executive explanation, evidence for each confirmed issue, affected assets, prerequisites, business impact, proposed fixes, coverage gaps and a retest. Separate a confirmed weakness from an untested hypothesis. A list of scanner findings without applicability review does not answer whether your business is exposed. Compare proposals on those deliverables; we have not verified service prices. ## Download the infrastructure self-test kit - [ARTEX-style attack self-test kit (.md)](/downloads/artex-style-attack-self-test-kit.md): authorization boundaries, an internet-facing service inventory, mapped configuration checks, read-only commands, dated detection markers, a monthly retest calendar and a hired-test scope. For an IT provider turning this review into a recurring client service, *[The $20 Dollar Agency](/books/)* pairs with the business side: define the deliverable, name the person responsible and include the follow-up that closes unresolved findings. ## The Computer Fraud and Abuse Act and other legal risks Running intrusion tools against systems you do not own or have permission to test can create liability under the Computer Fraud and Abuse Act, [18 U.S.C. § 1030](https://www.law.cornell.edu/uscode/text/18/1030). Owner authorization in writing establishes the boundary of a legitimate engagement; it does not grant rights over a supplier's other customers or excuse activity outside scope. DOJ's [May 2022 charging policy](https://www.justice.gov/archives/opa/pr/department-justice-announces-new-policy-charging-cases-under-computer-fraud-and-abuse-act) directs federal prosecutors not to charge qualifying good-faith security research. Its federal charging guidance does not grant immunity or permission to experiment on strangers' systems. State laws also apply, including [California Penal Code § 502](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=502.&lawCode=PEN) and [Texas Penal Code § 33.02](https://tcss.legis.texas.gov/resources/PE/htm/PE.33.htm). Exposure of customer information can trigger separate notification duties. The [FTC Safeguards Rule, 16 C.F.R. part 314](https://www.ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-business-needs-know), covers certain financial businesses, including some small nonbank businesses. Its notification amendment, effective in May 2024, requires covered institutions to notify the FTC as soon as possible and within 30 days after discovering a qualifying event involving unauthorized acquisition of unencrypted customer information affecting at least 500 consumers. Encrypted information also counts as unencrypted for this rule when an unauthorized person accessed its encryption key. States have their own notice laws. Ask counsel which duties and deadlines apply; the federal threshold is not permission to ignore smaller incidents. For more on permissions, contracts, data rights and contained work, read our articles on the [Cloudflare audit skill](/blog/cloudflare-security-audit-skill-small-business/), [rea](/blog/rea-reverse-engineering-agent-small-business-legal-risks/), [Scrapling](/blog/scrapling-web-data-small-business-legal-risks/), [Firecrawl](/blog/firecrawl-web-data-api-small-business/) and [MoneyPrinterTurbo](/blog/moneyprinterturbo-ai-short-videos-small-business/). ## Related reading - [ARTEX, Claude agents and South Korean banks: small business lessons](/blog/artex-claude-agents-south-korean-banks-small-business-lessons/) - [Cloudflare's security audit skill](/blog/cloudflare-security-audit-skill-small-business/) - [rea and reverse engineering legal risks](/blog/rea-reverse-engineering-agent-small-business-legal-risks/) - [Scrapling and permission for web data](/blog/scrapling-web-data-small-business-legal-risks/) - [Firecrawl for small businesses](/blog/firecrawl-web-data-api-small-business/) - [MoneyPrinterTurbo and AI video responsibilities](/blog/moneyprinterturbo-ai-short-videos-small-business/) - [GitHub's top repos of 2026 for small businesses](/blog/github-top-repos-2026-small-business/) - [AI security's three trust boundaries](/blog/ai-security-three-trust-boundaries/) ## Fact-check notes and sources Read dates below are November 3, 2026, UTC. Campaign figures and model labels are attributed vendor observations. We did not measure them on our infrastructure. Defensive recommendations and the review schedule are our interpretation. No paid product or service prices are quoted. - **Architecture, scope and controls:** [CrowdStrike VAPOR PANDA report](https://www.crowdstrike.com/en-us/blog/vapor-panda-uses-artex-to-scale-global-campaign/), 22:31:23: October 29 publication; activity from October 17; more than 200 targets, at least 14 countries, 11-day window, nine confirmed post-exploitation cases; shared graph, agent roles, human steering, approximate token consumption, weaknesses, markers and ATT&CK mappings. - **Earlier campaign and models:** [CrowdStrike South Korean finance report](https://www.crowdstrike.com/en-us/blog/unknown-threat-actor-uses-artex-to-target-south-korean-finance/), 22:31:22: October 7 publication, open directories, tool and model labels, attribution limits and indicators. No personal identifying details are reproduced. - **Cloud application trust:** [Microsoft Storm-0940 report](https://www.microsoft.com/en-us/security/blog/2026/10/21/storm-0940-uses-artex-agentic-ai-connectwise-supply-chain-compromise/), 22:31:22: October 21 publication, framework description, less-than-three-hour examples, more than 2,250 attempted tenant sign-ins, evidence limits and remediation guidance. - **Project status:** [Developer profile](https://github.com/Autumn-27), 22:31:22; [profile README lines 5 to 18](https://github.com/Autumn-27/Autumn-27/blob/main/README.md#L5-L18), 22:32:11. Translation is ours. The October 8 closure announcement is verified; original release day and exact repository removal time are not. - **Malware background:** [Elastic BLOODALCHEMY analysis](https://www.elastic.co/security-labs/threat-command/disclosing-the-bloodalchemy-backdoor), 22:31:22: October 13, 2023 publication. No ARTEX mention found in this page; the later association comes from CrowdStrike. - **AI provider position:** [Anthropic September report](https://www.anthropic.com/threat-intelligence-report-september-2026), 22:31:24; [usage policy](https://www.anthropic.com/legal/aup), 22:31:25. These establish general misuse response and authorization rules; neither provides an ARTEX-specific statement. - **Supporting reporting:** [BleepingComputer's global targeting report](https://www.bleepingcomputer.com/news/security/chinese-hackers-use-artex-ai-in-automated-attacks-against-200-orgs/), 22:31:22, published November 2; [former pentester report](https://www.bleepingcomputer.com/news/security/former-pentester-used-artex-ai-to-steal-82-million-data-records/), 22:31:24, published October 26. Police-related counts are attributed through its local-media reporting. - **Configuration reference:** [AWS instance metadata options](https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/configuring-instance-metadata-options.html). Read 22:31:30; command validation notes accompany the downloadable kit. - **Legal references:** [Federal statute](https://www.law.cornell.edu/uscode/text/18/1030), 22:31:25; [Texas chapter 33](https://tcss.legis.texas.gov/resources/PE/htm/PE.33.htm), 22:32:22. [DOJ charging policy](https://www.justice.gov/archives/opa/pr/department-justice-announces-new-policy-charging-cases-under-computer-fraud-and-abuse-act) and [California § 502](https://leginfo.legislature.ca.gov/faces/codes_displaySection.xhtml?sectionNum=502.&lawCode=PEN): supplied browser excerpts read before 21:39 UTC. [FTC Safeguards Rule guide](https://www.ftc.gov/business-guidance/resources/ftc-safeguards-rule-what-your-business-needs-know): supplied browser excerpt read before 22:33 UTC. Exact read seconds were not supplied for those browser excerpts. *This post is informational, not legal or security advice for your specific systems. Mentions of third parties are nominative fair use. No affiliation or endorsement is implied.* --- Canonical HTML: https://jwatte.com/blog/how-artex-agentic-pentest-framework-works-defense/ RSS: https://jwatte.com/feed.xml JSON Feed: https://jwatte.com/feed.json Hero image: https://jwatte.com/images/how-artex-agentic-pentest-framework-works-defense.webp