Agentic attack frameworks
Behaviour first. The addresses have already moved.
What to look for after the FBI’s advisory on ARTEX and the campaigns using it. The behaviours survive the infrastructure rotating and the addresses do not, so they come first here and the atoms come last — which is the order these documents usually recommend and never follow.
Before you act on any of this
Four things, in this order
- Look back to at least 1 October 2026. Sign-in logs — interactive, non-interactive and service-principal — directory API activity, cloud audit, web and WAF, DNS and proxy. Confirm the retention actually goes back that far before you conclude you found nothing.
- Vet before you block. Many of these are proxies, virtual private network exits or cloud hosts that get reassigned. The FBI’s own guidance on this list is to investigate before acting on it, and it draws no distinction between indicators — the same instruction covers every one it publishes. The Usecolumn below is our judgement about where each row is worth starting. Vet all of them.
- Treat clients and model endpoints as context, not evidence. A common HTTP client or a model API domain means nothing on its own. It means something beside one of the behaviours below.
- On a hit, preserve before you change anything. Export sign-in, audit and API records first. Disable rather than delete — deleting the application destroys the evidence of what it did. Then contain, rotate, and report to ic3.gov.
Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 21 October 2026
The shape of a run
What the detections are detecting
Five moves, each feeding the next — which is why a leaked key, a source map and a weak sign-up flow get chained inside an afternoon rather than filed as three findings.
Map
Enumerates the estate and tests what each credential it finds can actually reach — sub-second permission checks, test writes, deployment dry-runs — before committing to anything.
Enter
Takes whichever way in is cheapest: an unauthenticated API, an unsigned assertion, a credential in a front-end bundle, a partner's trust relationship.
Hold
Adds credentials to applications that already exist rather than creating new ones, because an app you already approved does not look like an intrusion.
Escalate
Harvests tickets, dumps credential stores, forges tokens against keys nobody rotated, and assigns itself the roles it is missing.
Take
Reads mail and object storage in bulk, stages it locally, and sends command output to rotating external endpoints.
Source: FBI, 13 November 2026
Detections
What to build, and what it will cost you
Each one carries the telemetry it needs, a threshold to start from, and — the part these packs normally leave out — what else in a healthy estate does exactly the same thing. Tune against your own baseline before you turn any of them on.
Machine-speed bursts
Compressed bursts of requests with small iterative variations, many failures followed by a success. Sub-agents duplicating or overwriting each other's requests is a tell the volume is not a human and not a stable script.
TelemetryWeb and WAF logs, API gateway, sign-in logs
Start at
Microsoft reported peaks above 2,900 sign-in attempts per minute, more than 200 sign-ins in under four minutes on four application credentials, and 48 accounts with over 1,000 failures each. Mandiant timed execution bursts at 2-25 seconds, one carrying 27 events in 23 seconds, and over 500 requests fuzzing filenames in a single sweep. Start an order of magnitude below your own peak legitimate rate.
What else fires this
Load tests, a retry storm after an outage, a mobile client with a broken token-refresh loop, and the first morning of a bulk migration. Baseline per application, not per tenant.
Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 21 October 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026
Sub-second capability testing
A credential being measured rather than used: permission checks in sub-second succession, test object writes immediately deleted, deployment validation dry-runs across many resource types, and dry-run flags on cloud API calls.
TelemetryCloud activity and audit logs, object-storage data plane
Start at
Measured in one reported run: a deployment-validation sweep across 11 resource types in 8 seconds, and test objects written and deleted 330-400ms apart. Start at any identity making validation-only calls across three or more resource types in a minute with no subsequent create; the sub-second write-then-delete is the higher-fidelity half.
What else fires this
Infrastructure-as-code plan runs, policy-compliance scanners, backup verification jobs and cost tooling all do this legitimately — which is why the rule needs an allowlist of your own automation identities before it is turned on, not after.
Sources: FBI, 13 November 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026
Bulk secret reads
One identity reading many secrets across several vaults in a short window — the step between finding a credential and using everything it unlocks.
TelemetryKey vault / secrets manager audit logs
Start at
Microsoft's published hunting logic: 20 or more successful secret reads across 3 or more vaults by a single identity within a 15-minute window. We have left their numbers rather than inventing our own.
What else fires this
Deployment pipelines at release time, secret-rotation jobs, and a newly onboarded cluster pulling its whole configuration. All three are predictable, so exclude them by identity and by schedule rather than by raising the threshold.
Sources: Microsoft Threat Intelligence, 21 October 2026; Microsoft Threat Intelligence, 5 November 2026 (cited, not read)
Credential probing
A failed client secret immediately followed by a success, or one source address authenticating as three or more different service principals.
TelemetryService-principal sign-in logs
Start at
Three or more distinct service principals from one address in an hour with at least one success. On Entra, invalid-secret errors AADSTS7000215 and AADSTS7000222 followed by a success is the sharper variant.
What else fires this
A shared egress address — a corporate gateway, a CI runner pool or a hosted integration platform — will legitimately present many service principals from one address. This detection is much better scoped to addresses outside your known automation ranges.
Sources: Microsoft Threat Intelligence, 5 November 2026 (cited, not read); Microsoft Threat Intelligence, 21 October 2026
Application persistence
A credential added to an application that already exists, a new owner, a new app-role assignment, or an application assigning itself a privileged role. This is the step that survives the original compromise being cleaned up.
TelemetryDirectory audit logs
Start at
Microsoft's logic flags an application adding itself to a high-privilege role with no prior self-assignment to that role in the preceding seven days. Credential additions to admin-consented applications deserve an alert at any volume.
What else fires this
Certificate rollover, a genuine ownership handover when someone leaves, and application registration automation. The first two are rare enough to review by hand; the third needs an allowlist.
Sources: Microsoft Threat Intelligence, 21 October 2026; FBI, 13 November 2026
Sequential enumeration
Incrementing identifiers, or a single session touching an unusually high number of records, against portals and lookup APIs.
TelemetryApplication and API logs
Start at
Per-session record counts above the 99th percentile for that endpoint, or any session where more than half the requested identifiers are sequential.
What else fires this
Reporting exports, a customer's own integration paging through their data, and support staff working a queue. Scope by account age and by whether the records belong to the caller — an account three minutes old enumerating is a different event from an account three years old doing it.
Sources: FBI, 13 November 2026; Zscaler ThreatLabz, 14 October 2026 (via FBI FLASH note 7; link reaches their index); Akamai, 23 October 2026 (via FBI FLASH note 1; link reaches their index)
New account, instant cross-tenant reach
An account created and confirmed, then within minutes requesting objects belonging to another tenant or another user.
TelemetryApplication and API logs
Start at
Any cross-tenant object request from an account less than an hour old. This one should be close to zero in a healthy system, which is what makes it worth alerting on rather than dashboarding.
What else fires this
Shared-workspace invitations and genuine multi-tenant customers, if your model allows one identity across tenants. If it does, scope to requests that are refused.
Sources: FBI, 13 November 2026; Akamai, 23 October 2026 (via FBI FLASH note 1; link reaches their index)
Server egress somewhere new
A production server reaching request-capture services, tunnelling providers, file-sharing or external model APIs it has never used before. Command output being posted to rotating external endpoints is the reported pattern.
TelemetryProxy, DNS and firewall logs
Start at
First-seen external destination for a given server, measured against a 30-day baseline. Model API endpoints reported in this activity include api.anthropic.com, api.deepseek.com, open.bigmodel.cn and api.x.ai — a production database server calling any of them is worth a look.
What else fires this
A legitimate new integration, a package manager reaching a new mirror, and any team that has genuinely started building with a model API. The point of this one is not that the destination is bad; it is that the source has no business going there.
Sources: FBI, 13 November 2026; Zscaler ThreatLabz, 14 October 2026 (via FBI FLASH note 7; link reaches their index)
Exposed workload turning inward
An internet-facing host initiating directory, authentication, file-sharing or remote-management protocols into the internal network. Kerberoasting and directory replication requests belong here.
TelemetryEndpoint detection, domain controller logs, internal network flow
Start at
Any occurrence. A host in a perimeter network has no legitimate reason to authenticate inward, so this is a rule rather than a threshold.
What else fires this
Poorly segmented legacy applications that genuinely do this, and backup agents. Both are findings in their own right, so the false positives here are worth the triage.
Source: FBI, 13 November 2026
Looking for your blind spots
Inventory queries that enumerate which hosts are missing an endpoint sensor. The agent is mapping where it will not be seen before it moves.
TelemetryEndpoint detection, script-block logging
Start at
Any query correlating host inventory against endpoint-agent presence from a non-administrative context.
What else fires this
Your own asset-management and compliance tooling, which does exactly this on a schedule. Allowlist it by identity and by host, and the remainder is high-fidelity.
Source: FBI, 13 November 2026
A feature turned into a megaphone
A notification or email feature that suddenly sends to hundreds of internal users — a legitimate function used as a mass-messaging tool.
TelemetryApplication logs, mail flow
Start at
Recipient count per send above your own 99th percentile for that feature, or any send above a fixed internal ceiling.
What else fires this
Genuine announcements, incident notifications and marketing sends. This is a cheap detection because the legitimate cases are scheduled and the illegitimate one is not.
Source: FBI, 13 November 2026
Web shell candidates
New server-executable files with short incremental names appearing in web application paths, and database logging reconfigured to write into a web-served directory.
TelemetryFile integrity monitoring, web server paths, database query logs
Start at
Any new executable file in a web root outside a deployment window. Any change to a database general-log path is worth alerting on by itself.
What else fires this
Deployments, and plugin or module installation in content-managed applications. Tie this to your release calendar — a new file in a web root during a deploy is noise, and the same file at 3am is not.
Sources: FBI, 13 November 2026; Zscaler ThreatLabz, 14 October 2026 (via FBI FLASH note 7; link reaches their index)
Queries
The logic, then one implementation of it
The code is for one vendor’s stack. The logic above each one is not, and it is the half you need if you are anywhere else — which is most people, and which is the half these packs usually omit while telling you to translate it.
Are the compromised vendor applications in my tenant, and what can they do?
Take the list of application identifiers. Join it against your directory's registered applications. For each match, return the display name, whether it was admin-consented, every permission granted, and the privilege level of each. Do this even where the integration is no longer in use — the grant outlives the usage.
Defender XDR advanced hunting
let VendorApps = dynamic([
"506e08f9-6e9a-4da4-942b-a099a03644ae",
"24205314-2b89-4e07-a2a4-579e898a5640",
"0de14c77-b9b0-4137-9360-34878a929827"
]);
OAuthAppInfo
| where OAuthAppId in (VendorApps)
| mv-expand Permission = Permissions
| project AppName, OAuthAppId, ServicePrincipalId, AppStatus,
PrivilegeLevel, IsAdminConsented,
PermissionName = tostring(Permission.PermissionName),
PermissionPrivilege = tostring(Permission.PrivilegeLevel)Before you run it: Search by application identifier, never by display name — a display name is attacker-controllable and is not unique across tenants.
Adapted from: Microsoft Threat Intelligence, 21 October 2026
Did anything sign in as a service principal, or call the directory API, from reported infrastructure?
Union service-principal sign-ins with directory API activity, normalise the address column across the two schemas, filter to the reported addresses, and order by time. Then widen: the same query with the address filter removed, grouped by source network, is the one that finds infrastructure nobody has published yet.
Defender XDR advanced hunting
let IPs = dynamic([
"144.172.64.60","144.172.64.99","144.172.64.128",
"187.15.88.186","187.15.90.217","187.15.89.186"
]);
union kind=outer
(EntraIdSpnSignInEvents | extend EventType = "SPN sign-in"),
(GraphAPIAuditEvents | extend EventType = "Graph API",
IPAddress = IpAddress)
| where Timestamp >= datetime(2026-10-01)
| where IPAddress in (IPs)
| order by Timestamp ascBefore you run it: The address column is `IPAddress` in one of these tables and `IpAddress` in the other. The casing difference is real and silently returns nothing if you do not normalise it, which is why the extend above exists.
Adapted from: Microsoft Threat Intelligence, 21 October 2026
Is one source address authenticating as several different service principals?
Group service-principal sign-ins by source address in hourly buckets. Count distinct service principals, successes and failures. Return buckets where one address used three or more distinct service principals and at least one attempt succeeded. This is behavioural: it keeps working after every published address has rotated.
Defender XDR advanced hunting
EntraIdSpnSignInEvents
| where Timestamp > ago(30d)
| summarize SPNs = dcount(ServicePrincipalId),
Failures = countif(ErrorCode != 0),
Successes = countif(ErrorCode == 0)
by IPAddress, bin(Timestamp, 1h)
| where SPNs >= 3 and Successes > 0Before you run it: Exclude your own egress ranges first or this returns your CI estate. Validate the column names against your own tenant before scheduling it — this is a behavioural variant rather than a published query.
Adapted from: Microsoft Threat Intelligence, 21 October 2026
Did one identity read a lot of secrets across several vaults at once?
Filter vault audit events to successful secret reads. Bucket by 15 minutes and by identity. Count the reads and the distinct vaults. Return buckets with at least 20 reads across at least 3 vaults, and keep the source addresses for the follow-up.
Defender XDR advanced hunting
CloudKeyVaultEvents
| where Timestamp > ago(14d)
| where ActivityName == "SecretGet" and HttpStatusCode == 200
| summarize SecretGets = count(),
VaultCount = dcount(AzureResourceId),
IPs = make_set(IPAddress, 20)
by bin(Timestamp, 15m), AccountObjectId, AccountTenantId
| where SecretGets >= 20 and VaultCount >= 3Before you run it: Run it once with the thresholds removed to see your own deployment pipelines, then exclude those identities. Raising the threshold instead hides the attack as well as the pipeline.
Adapted from: Microsoft Threat Intelligence, 21 October 2026
Has anything on my network talked to reported infrastructure in the last 30 days?
Match outbound sessions against the address list, and summarise first seen, last seen and event count per source and destination pair. Vet before blocking: these are historical observations, and cloud addresses get reassigned.
Normalised network sessions (ASIM)
let ioc_ip_addr = dynamic([
"199.19.95.66","199.19.95.67","199.19.95.68",
"91.92.240.238","85.11.168.45","178.16.54.129",
"31.77.229.35","85.11.167.52"
]);
_Im_NetworkSession(starttime=ago(30d), endtime=now())
| where DstIpAddr in (ioc_ip_addr)
| summarize FirstSeen = min(TimeGenerated),
LastSeen = max(TimeGenerated),
EventCount = count()
by SrcIpAddr, DstIpAddr, DstDomain, DvcBefore you run it: The address list here is corrected — see the corrections below. The published sample query for this campaign contains 178.15.54.129, which is not the address in the indicator table it was written from.
Adapted from: Microsoft Threat Intelligence, 6 November 2026; FBI, 13 November 2026
Indicators
The atoms, dated
Last on the page on purpose. Every row carries its campaign, its date, the report it came from, and where we think it is worth starting.
| Indicator | Type | Campaign | Context | Source | Use |
|---|---|---|---|---|---|
| 199.19.95[.]66199.19.95[.]67199.19.95[.]68 | IP address | Storm-3358, financial sector | Remote-access tool command and control6 November 2026 | FBI, Microsoft Threat Intelligence | Block |
| 91.92.240[.]23885.11.168[.]45178.16.54[.]12931.77.229[.]35 | IP address | Storm-3358, financial sector | Framework servers6 November 2026 | FBI, Microsoft Threat Intelligence | Block |
| 85.11.167[.]52 | IP address | Storm-3358, financial sector | Command and control6 November 2026 | FBI, Microsoft Threat Intelligence | Block |
| 144.172.64[.]60144.172.64[.]99144.172.64[.]128 | IP address | Storm-0940, vendor supply chain | Password spraying, then sign-ins and directory API activity inside customer environmentsReported 21 October 2026 | FBI, Microsoft Threat Intelligence | Block |
| 187.15.88[.]186187.15.90[.]217187.15.89[.]186 | IP address | Storm-0940, vendor supply chain | Directory API access via compromised applications; transient addressesRelevant 15–16 October 2026 only | FBI, Microsoft Threat Intelligence | Hunt |
| 38.244.50[.]120 | IP address | South Korean financial institutions | Framework host with an open directoryReported 7 October 2026 | FBI, CrowdStrike, Censys | Hunt |
| 101.53.80[.]20205.214.59[.]31124.155.252[.]63154.201.79[.]24623.248.249[.]9023.158.220[.]98103.248.148[.]84203.160.133[.]172209.209.85[.]38 | IP address | South Korean financial institutions | Operator proxiesReported 7 October 2026 | FBI, CrowdStrike | Hunt |
| xcai[.]pro | Domain | South Korean financial institutions | Model API reseller used by the operatorReported 7 October 2026 | FBI, CrowdStrike | Hunt |
| 506e08f9-6e9a-4da4-942b-a099a03644ae | Application ID | Storm-0940, vendor supply chain | Vendor cloud backup applicationReported 21 October 2026 | FBI, Microsoft Threat Intelligence | Hunt |
| 24205314-2b89-4e07-a2a4-579e898a5640 | Application ID | Storm-0940, vendor supply chain | Vendor collaboration integrationReported 21 October 2026 | FBI, Microsoft Threat Intelligence | Hunt |
| 0de14c77-b9b0-4137-9360-34878a929827 | Application ID | Storm-0940, vendor supply chain | Vendor directory API integrationReported 21 October 2026 | FBI, Microsoft Threat Intelligence | Hunt |
| azure-toolkit/1.0 | User agent | Storm-0940, vendor supply chain | Sign-ins and directory API calls. The most distinctive of the set.Reported 21 October 2026 | Microsoft Threat Intelligence | Hunt |
| Python-urllib/3.12python-requests/2.31.0python-requests/2.34.2curl/7.88.1 | User agent | Storm-0940, vendor supply chain | Sign-ins, directory API and web probing. Common clients — these mean nothing on their own.Reported 21 October 2026 | Microsoft Threat Intelligence | Context |
| api.anthropic.comapi.deepseek.comopen.bigmodel.cnapi.x.ai | Domain | Multiple | Model API endpoints. Legitimate for a developer workstation; worth a question from a production database server.Reported 13 November 2026 | FBI, Zscaler ThreatLabz | Context |
| *.aspx, *.jsp in unexpected web application paths | Artefact | Multiple | Web shellsReported 13 November 2026 | FBI, Zscaler ThreatLabz | Hunt |
| artex-probe-NNNNN (five digits, incrementing) | Artefact | Multiple | Access-probe files left in web roots, storage buckets and file shares. High fidelity — the name is the tool's own.Reported 30 October 2026 | Google Threat Intelligence Group and Mandiant | Hunt |
| Accounts containing ArtexProbe | Artefact | Multiple | Persistence accounts in directories and application user tables, often with incrementing numbers and a plaintext password stored beside them.Reported 30 October 2026 | Google Threat Intelligence Group and Mandiant | Hunt |
| cl1217.jspf1219a.jspz1219.jsp | Artefact | Multiple | Web shell naming pattern rather than fixed names: short alphanumeric plus date-like digits, in a web-served path. Hunt the shape, not these three strings.Reported 30 October 2026 | Google Threat Intelligence Group and Mandiant | Hunt |
| SET GLOBAL general_log_file -> a web-served path | Artefact | Multiple | Database query logging repointed into a web directory so the log itself becomes a web shell. Any change to this setting is worth an alert on its own.Reported 30 October 2026 | Google Threat Intelligence Group and Mandiant | Hunt |
| webhook[.]site | Domain | Multiple | Legitimate request-capture service used for command output; one run rotated through 491 URLs on it. Alert on server egress, not on the domain alone.Reported 30 October 2026 | Google Threat Intelligence Group and Mandiant, FBI | Hunt |
| Self-hosted model gateway or reseller software on your own hosts | Artefact | Multiple | The FBI asks for evidence of self-hosted model gateways and unfamiliar model resellers in egress logs as a reportable artefact.Reported 13 November 2026 | FBI | Hunt |
Block — after vetting it against your own telemetry, per step 2 above; several of these are proxies or cloud hosts that get reassigned. Hunt — search for it, do not block it. Context — meaningless on its own, and worth something only beside one of the behaviours above.
Addresses re-armed for ingestion. Built from the same table above, so the two cannot disagree. Reviewed 18 November 2026.
Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 21 October 2026; Microsoft Threat Intelligence, 6 November 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026; CrowdStrike, 7 October 2026 (via FBI FLASH note 6; link reaches their index); Zscaler ThreatLabz, 14 October 2026 (via FBI FLASH note 7; link reaches their index); Censys, 5 November 2026 (via FBI FLASH note 4; link reaches their index)
Corrections
What we are carrying
Three things we found wrong in the published material, and one trap in the platform. Each was found by reading a query against the table it was written from, and each is checkable from the source beside it.
An address in a published sample query does not match its own indicator table
The sample network-session query for the financial-sector campaign contains 178.15.54.129. The indicator table in the same post, and the FBI's table, both give 178.16.54[.]129 — the third octet differs. Use 178.16.54[.]129. Anyone who pasted the query without reading the table has been hunting an address that was never reported.
Checked against: Microsoft Threat Intelligence, 6 November 2026; FBI, 13 November 2026
A vendor advisory says four application identifiers and lists three
The supply-chain post instructs defenders to inventory all four application identifiers listed in the profile; three appear in the published content. We publish the three that exist. If you are working from this list, treat it as incomplete and search by publisher rather than by identifier alone.
Checked against: Microsoft Threat Intelligence, 21 October 2026
The address column changes capitalisation between two tables in the same query
Service-principal sign-in events expose the source address as IPAddress; directory API audit events expose it as IpAddress. A union across the two filtered on either spelling silently returns half the data and no error. The query above normalises it.
Checked against: Microsoft Threat Intelligence, 21 October 2026
The other half: what to change
A 72-hour and 30-day matrix, and seven control areas with a test for each one that can come back negative.
Hunting tells you whether it already happened. Testing tells you whether it can. How our engine works, or point it at your external surface.