On this pageWhat an Agentic Attack Framework IsARTEX as the Documented ExampleParts of a FrameworkThe Operator Console and InfrastructureHow a Run UnfoldsHow Agent Activity DiffersLog Signals of Agent ActivityTraces After Initial AccessEvidence and Baselines to PrepareResponding to a Suspected RunWhat Defenders ChangeService Providers and Multi-Tenant ApplicationsMeasuring Detection ReadinessThe Same Design in Authorized TestingAgentic Attack Frameworks and ThreatDefenceFrequently Asked Questions An operator sets a target and a goal, and the agents plan the steps, run the tools and record the results with little further input. The best documented example, ARTEX, was built for authorized penetration testing and has been used in real intrusions since at least August 2026. This guide explains how these frameworks are built, how a run unfolds against an organization, what the activity looks like in sign-in, web and cloud logs, which records to prepare, how to respond to a suspected run, what defenders change, and how authorized testing uses the same design. What an Agentic Attack Framework IsAn agentic attack framework is an orchestration layer that connects an AI model to ordinary attack tools and keeps the model working toward a goal across many steps. It typically contains no exploit of its own. The techniques are familiar ones, such as password spraying, token forgery and cloud enumeration, and the framework contributes speed and persistence. A scanner or an attack script runs a prepared sequence, and an operator with an AI assistant still chooses each step in person. In an agentic framework the agents choose the next step from what the earlier steps returned, so the run adapts to each target without waiting for a person. The framework runs on the operator’s own server and reaches its targets over the network, so a victim system typically holds only the traces of what the agents did. ARTEX as the Documented ExampleARTEX is the agentic framework that government and Microsoft reporting describes in most detail, and this guide draws on it. Microsoft describes it as running a test “from reconnaissance through data exfiltration with little needed human interaction”. Origin: ARTEX was published as open source in July 2026 for authorized testing. Status: Its developer made it closed source on October 8, 2026, and the repository has since been taken down. Archived copies, renamed builds and a commercial version still circulate. Malicious use: The FBI and Microsoft have reported state-aligned and criminal groups using it in real intrusions. The guide to AI hacking covers those cases and the wider use of AI by attackers. Exposure: Operators run ARTEX as a hosted web console, and the FBI’s FLASH of November 13, 2026 puts the number of consoles exposed to the internet above 1,600. Some of those consoles belong to authorized testers. Parts of a FrameworkA framework of this kind is built from a task definition, a planning agent, worker agents, a shared memory, a set of tools and a connection to one or more AI models. ARTEX arranges them as follows. Task: The operator defines the scope, any accounts to use and the goal, such as gaining administrator access. Each task carries allow and deny rules, a timeout and a wrap-up step. Planner agent: The planner turns the goal into individual attack intents and hands them out. Worker agents: Workers claim intents, run the tools, several at a time, and write the results back. Shared memory: An asset graph records what the target looks like and an exploration graph records what has been tested. Facts, findings and hints sit on a shared board that every agent reads. Tool use: The agents run standard tools such as nmap, curl and Impacket, a recording proxy keeps the requests and responses, and results from a companion discovery platform load straight into the asset graph. Model failover: The framework has no model of its own and accepts commercial, reseller and self-hosted models. When one provider cuts off a key it moves to the next, so a campaign can be spread across several providers. The shared memory is the part that separates this design from a script. Every later step starts from what earlier steps recorded, which is how a route name found in one file becomes a request a few minutes later. A single fact can be followed from the worker that reads it to the worker that uses it. A worker reads a file. A worker downloads a source map from the target’s web application and finds the name of an internal API route in it. The fact is recorded. The worker adds the route to the asset graph and posts it to the shared board. The planner reacts. The planner reads the board and issues new intents against that route, such as calling it with no session or asking it for another tenant’s data. Workers claim the intents. Free workers send the requests, and the recording proxy keeps each request and response. The result is marked. The exploration graph notes each test as done, and a response that returns data is stored as a finding. The finding becomes a starting point. A credential or an internal address in that response goes back onto the board as a hint, and the planner plans from it. The loop explains why weaknesses on separate systems end up connected, since what one worker reads on a web application may be used minutes later by another at a sign-in page or a cloud API. What the agents do after they obtain credentials depends on the model behind them, so two runs may differ in detail. The Operator Console and InfrastructureThe operator works through a web console on a server they control, where tasks are entered, the graphs and findings are viewed and the model connections are set. In ARTEX the console is a web application backed by a database, and it listens on port 8787 by default. Console pages: The console presents tasks, pending intents, the exploration graph, findings, model profiles and intercepted traffic. Discovery feed: A companion asset discovery platform from the same developer supplies domains, subdomains, addresses, ports and endpoints, so a task may begin with the target already mapped. Model pool: The operator registers several model profiles, and the framework speaks the API formats common to commercial providers, resellers and self-hosted gateways. Renamed builds: The same platform runs under other names, including ScopeWeaver and LUCK, and a commercial edition is sold. A renamed build behaves like the original, so the tool’s name is a weak thing to detect on, and the behaviors described later in this guide apply to any framework built this way. For a defender, the console matters in the organization’s own address space and in its outbound traffic. A console found on a cloud instance the organization pays for should match an authorized test record, and external attack surface management is the practice that finds such hosts. In outbound traffic, a server that reaches an unfamiliar AI proxy or model reseller deserves a look. How a Run UnfoldsOnce the operator has set the task, a run typically moves from reading the exposed applications, to sign-in, to whatever the first working credential can reach. The steps below repeat across the documented ARTEX intrusions. Read the application. Within minutes of first contact the agent downloads JavaScript bundles, source maps, runtime configuration and API documentation, then calls internal API routes by the names it found there. Learn the baseline. Before each round of guessing, it requests a path that should not exist, to learn what a “not found” response looks like on that host. Probe sign-in. It sends OAuth callbacks with made-up codes and broken state values, and tries tokens that carry no signature. Spray passwords. Sign-in attempts arrive in long bursts at a fixed rhythm, with a new user agent for each burst. Abuse recovery and sign-up. The agent checks which usernames exist and works the reset and activation endpoints. Where sign-up is open, it registers an account and at once requests another tenant’s data. Use what it gets. With stolen application credentials, agents run by Storm-0940, a state-aligned actor that Microsoft tracks, signed in as the application, explored Microsoft Graph and added credentials to applications for persistence. With a stolen cloud access key, an agent searches logs for secrets and internal addresses. Enumerate what it can reach. The agent tries nearly every read-only discovery call available to the identity, and Microsoft saw bulk secret reads across several Azure Key Vaults. Adapt and continue. When a sign-in page starts rate limiting, the agent moves to another endpoint on the same host. When an address is blocked, a new one appears within minutes, and a run can continue for many hours. Sign-in is one route among several. In activity Microsoft attributes to Storm-3358, a financially motivated group, the framework and conventional scanners found anonymous file upload and server-side request forgery flaws in web applications, and the agents then used netcat and curl to install remote access tools. A person typically remains involved, and the operator may step in at decision points while the agents work several organizations at once. How Agent Activity DiffersAgent activity differs from a human operator’s in its timing, its repetition and the labels it leaves on its own work, and the automation that gives it speed also makes it noisy. Bursts and pauses: Inside a network, activity arrives in bursts of a few seconds separated by pauses while the model decides what to do next. Trial and error: Agents retry failed attempts with small changes, test things an experienced person would skip, and may overwrite each other’s work. Self-labeling: Test files, accounts and cloud roles often carry the tool’s name and a number that counts up. Overstated results: An agent may record success after a check that showed only that a host was reachable, so its own account of a run can claim more than it achieved. The activity is noisy enough to raise alerts in most environments, and the constraint is the time left to act on them. Log Signals of Agent ActivityIn sign-in, web and cloud logs, an agent run shows as regular timing, placeholder values and broad sweeps that an administrator would have no reason to produce. Signal Where it shows Why it stands out Failed sign-ins for many accounts from one source, evenly spaced at an interval of seconds Identity provider sign-in logs People and most spray tools do not hold a fixed rhythm for long. A request for a placeholder path such as /zzz-nonexistent-xyz123 right before a burst of guesses Web server, WAF and storage access logs A path that cannot exist has no legitimate caller. Internal API routes called minutes after source maps are downloaded Web server and WAF logs The route names are published nowhere else. The same unsigned token sent to many single sign-on paths Web server and WAF logs A token whose header declares alg: none has no valid use. User agents that name the tool, such as artex-pentest/1.0 Sign-in, API, cloud audit and web logs Operators do not always change the defaults. An application adding credentials or assigning itself a role Directory audit log Applications rarely change their own access. A sweep of read-only discovery calls that mostly fail Cloud audit log A new identity trying every call produces a spike of access denied errors. A burst of dry-run permission checks across many actions Cloud audit log An administrator tests one action, and the agent tests every action within seconds. A blob written and deleted within a second Cloud storage logs The write has no purpose other than to learn that writing works. Source addresses and user agents change quickly, and the FBI advises checking published indicators against current telemetry before blocking, because cloud addresses are reassigned. The behaviors in the table hold after the indicators rotate. Traces After Initial AccessOnce the agents are inside, the traces move from sign-in and web logs to endpoint, directory and cloud audit records. Stage What the agents do Where it is recorded Commands on web hosts A web or database service starts a shell, curl and discovery commands in quick succession. Endpoint process events, application service audit and HTTP logs Web shells Script files appear in web directories, in one case written by pointing a database log at a web-accessible file. File creation events on web servers, database audit logs Callbacks Command output is posted to a public request-collection service across rotating URLs. Process, proxy and DNS logs Directory attacks Kerberoasting, DCSync and reads of local administrator passwords follow a large directory enumeration. Domain controller security events Lateral movement A scan for remote management ports is followed by several remote execution tools tried in turn. Network sensors, logon events, endpoint detection Coverage checks Process listings run across the domain to find machines with no endpoint sensor. Process events on the machines that have one Cloud foothold A new SSH key pair, a dedicated public address and small instances used as jump hosts are created. Cloud audit log A server running a framework also calls an AI model provider. The FBI lists model provider endpoints as destinations to investigate when a server calls them unexpectedly alongside other suspicious activity, and many legitimate applications call the same endpoints. Evidence and Baselines to PrepareSeveral records that show an agent run are switched off or short-lived by default, and several signals become usable only once their legitimate causes are known. Record What it shows of a run Default to check Identity provider sign-in and audit logs The spray, its rhythm, any sign-in that succeeded, and credential or role changes on applications. Microsoft Entra ID keeps both logs for 30 days with a P1 or P2 license and seven days on the free tier, so longer history depends on export. Microsoft Graph activity logs Each request an application or user sent to Microsoft Graph, including the reads an agent makes while exploring. Collection starts only when the log category is enabled in diagnostic settings and routed to storage or an analytics tool. Cloud storage read and write logs Anonymous reads, placeholder object names and write-then-delete tests. Azure Storage resource logs are collected only once a diagnostic setting exists. Web server and WAF logs Source map downloads, placeholder paths, unsigned tokens and calls to internal routes. The full request path and user agent need to reach central storage before the host rotates its logs. Directory object access events Replication requests and large directory reads on domain controllers. Windows writes event 4662 only where directory service access auditing is enabled and the object carries an audit entry. Endpoint process and file events on servers Shells started by web and database services, and script files written into web directories. The events exist only on machines that carry a sensor. Retention sets how far back a hunt can reach. The FBI dates malicious use of ARTEX to at least August 2026, so an organization checking for earlier activity needs records from that month onward. A behavioral signal also needs a list of the legitimate activity that resembles it, written by the team that owns the system. Automation accounts: The identity team lists the applications and pipelines expected to add credentials or assign roles, so the same change made by any other application stands out. Deployment tooling: Server owners name the processes that normally write script files into web directories, which leaves a file written by a web or database service as the exception. Backup and deployment jobs: The cloud team names the callers that write and delete test blobs in normal operation, and a review starts with callers nobody can name. Directory synchronization: Accounts that replicate the directory by design are recorded, so replication requested by any other account is investigated at once. Responding to a Suspected RunA suspected agent run is handled like any intrusion, in an order set by its speed: preserve the records, establish how far the agents got, then contain by identity, key and host. Preserve the logs. Export sign-in, audit, API, cloud and web logs for the period before retention removes them. Confirm the pattern. Several signals from one source against one host within minutes, such as a source map download, a placeholder path and an unsigned token, make an agent the likely explanation. Establish how far it got. Failed attempts are reconnaissance, and a successful sign-in, an issued token or a first successful API call marks the point where the activity becomes an incident. Contain identities. Reset the affected accounts, revoke their sessions and review the MFA methods registered on them. For applications, remove credentials nobody can account for, then review role assignments, consent grants and owners. Contain cloud keys. Deactivate and rotate an exposed access key, review everything it did, and treat any secret that appeared in a log it searched as exposed. Check applications and servers. Remove accounts the agent registered, check whether any request for another tenant’s data was answered, and look on servers for web shells, new scheduled tasks, remote access tools and tunnels. Block and keep watching. Block the source addresses and user agents, expect replacements within minutes, and leave the behavioral detections running. Report. The FBI asks organizations to report such activity to a field office or to IC3 at ic3.gov, with console artifacts, sign-in logs and user agents where available. Where the agents reached a domain controller the response widens, and the krbtgt key and every service account password are treated as exposed. The ARTEX advisory lists the published indicators and gives the hunt for each log source. What Defenders ChangeDefenders mostly change priorities, because an agent finds existing weaknesses faster and tests every one of them. The steps below run from what the agent reads first to what limits it after access. Remove what the agent reads first. Take source maps and runtime configuration off production, and stop showing stack traces, SQL errors and internal paths to visitors. Harden every sign-in path. Reject unsigned tokens, validate OAuth state on every callback, and rate limit by account as well as by address across all sign-in and recovery endpoints on a host. Protect the targeted accounts. Require phishing-resistant MFA for administrators and engineers, and close self-registration where it is unused. Shorten the life of credentials. Rotate or remove long-term cloud access keys, keep secrets out of logs, and apply least privilege to multi-tenant applications. Keep administration interfaces off the internet. That includes application publishing endpoints, cloud consoles, CI/CD systems, hypervisors and backup platforms. Limit what a server can do. Block outbound calls to request-collection and tunneling services, and put an endpoint sensor on every server. Record the evidence. Send Microsoft Graph activity logs and cloud storage data events to the SIEM, since those records are otherwise never written. Detect behavior and plant decoys. Keep rules for fixed-rhythm sprays, self-granted application access and enumeration bursts, and place decoy credentials and honeypots where an agent would look. Service Providers and Multi-Tenant ApplicationsService providers and software vendors are targeted because one of their application credentials can open many customer environments. In activity Microsoft attributes to Storm-0940, agents signed in to customers’ Microsoft 365 tenants through a cloud backup vendor’s compromised applications. Other documented runs went at IT and software vendors directly, working their sign-in pages, web applications and Microsoft 365 integrations to reach customer data. Inventory: The provider lists every application it operates that holds consent in client tenants, and each client lists the vendor applications it has consented to, including ones it has stopped using. Least privilege: Permissions on a multi-tenant application are cut to what the service needs, since an agent explores whatever the application can reach. Records in the client tenant: Microsoft Graph activity logs are collected by the tenant that holds the data, and Microsoft notes that they do not show a multi-tenant application’s activity in another tenant. Each client tenant needs the logs enabled for the provider’s application to be traceable there. Measuring Detection ReadinessReadiness for an agent run is measured by what the organization can see, how quickly it acts and how much it leaves exposed. Signal coverage: The share of the log signals in this guide that have a detection confirmed by a test, as distinct from one assumed to work. Sensor coverage: The proportion of servers that carry an endpoint sensor, including servers with no internet exposure. Time to triage: The time from the first signal of a run to a decision by a person or an automated rule. Time to containment: The time from a first successful sign-in or API call to session revocation or key deactivation. Exposure counts: The number of long-term cloud access keys, of applications holding tenant-wide consent, and of production hosts that still serve source maps. Microsoft has reported runs that reached privileged cloud access within hours of first targeting, which sets the scale for the time measures. An authorized agent run gives real values for them, and the guide to purple teaming covers how the detection gaps from an exercise are recorded and closed. The Same Design in Authorized TestingAuthorized agentic penetration testing uses the same parts: a task, a planner, workers and a shared record of what has been learned. ARTEX itself was written for that purpose, and its task definition already carries scope rules. The difference lies in who holds the controls and where they are enforced. Some ARTEX versions ship with a default blocklist of government and education domains, and the operator can change it, so the limits bind only an operator who wants them. The guide to AI penetration testing explains the controls an authorized run adds and how they are enforced outside the AI model. An authorized agent run produces many of the log signals described above, so a monitored organization can use a test to check which of them its own detections catch. Agentic Attack Frameworks and ThreatDefenceThe planner and worker design described in this guide is also how ThreatDefence runs authorized testing. A Planner agent coordinates the work, a testing agent tests a customer’s exposed systems, and the limits are agreed with the customer before anything runs. Agreed limits: Tests cover confirmed and authorized scope only, inside agreed time windows and capped request rates, with denial of service, data modification and data removal excluded. Enforcement: The agents work in isolated sandboxes whose network policy enforces the scope, and higher-impact tests wait for approval. Record: Every request and result is logged, and each finding carries its proof. Hunting: For a monitored customer, the Defender and Hunter agents go through telemetry for signs that a confirmed weakness was already used. The testing sits inside an AI SOC platform with a 24x7 SOC behind it, so SOC analysts review significant findings alongside the monitoring, compliance and cyber risk work the platform automates. Agentic penetration testing describes the service. Frequently Asked QuestionsIs an agentic attack framework malware?No. It is software the operator runs on their own infrastructure, so there is typically no framework file to find on a victim system. What the agents leave behind is evidence of their work: web shells, test files and accounts that carry the tool’s name, new cloud resources and remote access tools.What is ARTEX?ARTEX is an open-source penetration testing framework, published in July 2026, in which a planner agent and worker agents carry out a test toward a goal the operator sets. It was written for authorized testing, and the FBI dates its malicious use to at least August 2026. The project went closed source in October 2026, and copies still circulate.Does blocking an AI model provider stop an agentic attack?No. The framework is built to fail over between providers, and an operator can run a self-hosted model that no provider controls. The model traffic also runs between the operator’s server and the provider, outside the target’s network, so the target organization’s controls belong on its own sign-in pages, applications and cloud accounts.Do decoys work against attack agents?Yes, according to Microsoft, which reports that honeypots, decoy credentials and tarpits draw agents away from real systems and produce high-confidence alerts. An agent tends to follow a convincing lead where a person might hesitate. Decoys work as an addition to the controls on sign-in and exposure, since they detect an agent that is already probing.Will rate limiting stop an agent?Partly. A limit slows the agent on the page where it applies, and in the documented runs the agent then shifted to other paths or came back from fresh addresses. A limit keyed to the account and shared across all sign-in and recovery paths on a host leaves it far less room to continue.Which logs record agentic attack activity?Identity provider sign-in and audit logs, web server and WAF logs, and cloud audit logs record the approach. Endpoint process and file events, domain controller security events and network sensors record what follows access. Two of the most useful sources are often switched off, Microsoft Graph activity logs and cloud storage data events, and are worth enabling before they are needed.Are there agentic attack frameworks other than ARTEX?Yes. The same platform circulates under other names, including ScopeWeaver and LUCK, and as a commercial edition, and other AI attack frameworks have been found on servers that also host ARTEX consoles. The planner, worker and shared memory design is general, which is why detection built on behavior carries over from one framework to the next.Is every detected agent run an attack?No. Authorized testers use the same tooling, and some exposed consoles belong to them. An organization separates the two by its own records: a signed authorization, a test window and the source addresses given in advance. Activity that matches none of those is treated as hostile until shown otherwise.Is an agentic attack framework the same as an AI penetration testing tool?Yes, in design. Both use a task, a planning agent, worker agents and a shared record, and ARTEX was written for authorized testing. An authorized test adds written permission, a scope enforced outside the AI model, excluded methods, approval for higher-impact steps and a record the system owner can read.