Skip to content

FBI FLASH-20261113-001

Defending against agentic attack frameworks

Attackers are running intrusions with open-source agentic frameworks — ARTEX and its forks — and the FBI reports some of them reaching privileged access in under three hours.

The published reporting on that activity, turned into actions. We have not observed it: every block cites the report it came from and says whether we read it.

The next 72 hours, and the 30 days after

Priority actions · agentic attack frameworks

Reviewed 18 November 2026

Area

Next 72 hours

Next 30 days

Self-service and APIs

Next 72 hoursRate-limit registration, login, password reset and lookup. Require a verified address before a new account can read anything.

Next 30 daysInventory every public endpoint, including the legacy ones, and run authorisation tests for cross-tenant access.

Machine identities

Next 72 hoursScan repositories, CI and images for secrets, including git history. Rotate what you find — removing it from code does not revoke it.

Next 30 daysReplace long-lived keys with short-lived, workload-bound credentials. Give every one an owner, a purpose and an expiry.

Exposed surface

Next 72 hoursTake admin consoles, deployment and publishing interfaces, and backup platforms off the public internet.

Next 30 daysMove internal and partner applications behind brokered access. Require architecture review before anything new gets a public address.

Vendor trust

Next 72 hoursReview consented applications with tenant access by application ID, not display name. Remove the stale ones.

Next 30 daysBuild a register of which third party holds which credential, at what scope, and who owns the relationship.

Money movement

Next 72 hoursRequire out-of-band callback on any new or changed payee, to a number already on file — never one in the request.

Next 30 daysAdd dual approval and thresholds to payment workflows, and write down that a voice or video call authorises nothing.

Detection and response

Next 72 hoursSearch back to at least 1 October 2026 for the behaviours on the hunting page. Confirm your sign-in, audit and Graph telemetry is actually retained.

Next 30 daysStand up the behavioural detections and pre-approve containment for the high-confidence ones.

Testing

Next 72 hoursPoint an authorised test at your own external surface — the whole of it, not the flagship product.

Next 30 daysMake that test part of the release process rather than an annual event.

Each row maps to a section below, where the reasoning and the test for it live. The left column is this week.

The seven areas

Each area, in full

The matrix gives one action per area for this week and one for this month. Below is each of those same seven areas in full: why attacks land there, a test you can run to find out whether yours is exposed, and everything we would do about it.

Self-service and APIs

Why here

Anything a stranger can do without talking to a salesperson will be done thousands of times, in parallel, by something that does not get bored.

The test

Register on your own product with a disposable address. Before any human approves anything, try to read, list or enumerate something belonging to another tenant. Time how long the whole attempt took you.

Passing looks like: Every attempt is refused server-side, the account cannot reach another tenant's identifiers at all, and the attempt itself raised an alert. If it was merely slow, it was not refused.

What to do

  • Require a verified email or phone before a new account can read data, call APIs or invite users, and block disposable domains.
  • Rate-limit registration, login, password reset, username lookup, activation and token resend per address, per network and per device fingerprint — and return identical responses for valid and invalid usernames.
  • Enforce authorisation server-side on every object lookup. Test that tenant A cannot read tenant B's records by changing an identifier.
  • Return only the fields the caller needs. Never return stack traces, SQL errors or framework details to an unauthenticated caller — an agent reads errors to correct itself.
  • Treat a new account that immediately probes another tenant's resources as an incident, not as a 403.
  • Require strong re-authentication for high-risk password resets and account recovery, and alert on bursts of resets, account creations and credential additions — the recovery flow is the way in that does not need a password at all.

Sources: FBI, 13 November 2026; Akamai, 23 October 2026 (via FBI FLASH note 1; link reaches their index); Zscaler ThreatLabz, 14 October 2026 (via FBI FLASH note 7; link reaches their index); Google Threat Intelligence Group and Mandiant, 30 October 2026

Machine identities

Why here

A long-lived key that works from anywhere turns one leak into general access. Microsoft timed one agent from a secret found in public code to data out of a cloud vault at about five minutes.

The test

Pick a service principal, service account or deployment token at random. Say who owns it, what it is for, and when its credential expires.

Passing looks like: All three answers inside a minute, from a system rather than from memory. If the expiry is 'it does not', the inventory is a list of names.

What to do

  • Give every service principal, service account, OAuth application, API key and deployment token an owner, a purpose and an expiry.
  • Go short-lived: workload identity federation and managed identities instead of static client secrets and access keys.
  • Constrain where, not just what — restrict workload identities to expected networks and workloads, and restrict role assumption to named roles.
  • Scan continuously: repositories, git history, container images, build artefacts and CI logs. Rotate anything found — deleting it from code does not revoke it.
  • Put secret stores behind private paths, alert on bulk reads, and baseline normal automation first so the alert means something.

Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 5 November 2026 (cited, not read); Microsoft Threat Intelligence, 21 October 2026

Exposed surface

Why here

It cannot attack what it cannot reach, and what it reaches first is whatever gets less attention than the flagship product: partner portals, mobile backends, employee apps, legacy services.

The test

From an address outside your network, with no credentials, list everything of yours that answers. Then find the one nobody on the call recognises.

Passing looks like: No admin console, deployment or publishing interface, hypervisor, backup platform or certificate service answers at all. The list contains nothing you cannot name an owner for.

What to do

  • Inventory every internet-facing application, including the peripheral ones: partner and supplier portals, mobile backends, employee apps, legacy services and third-party integrations.
  • Restrict cloud administration, deployment and publishing interfaces, hypervisors, backup platforms and certificate services to trusted networks or dedicated administrative paths.
  • Require phishing-resistant multi-factor authentication for administrators, engineers and finance first.
  • Stop internet-facing workloads initiating directory, file-sharing or remote-management protocols into the internal network, and lock down egress so a compromised host cannot reach tunnelling services or external model APIs.
  • Harden the identity infrastructure the agent pivots through: Credential Guard, Protected Users, SMB and LDAP signing with channel binding, reduced NTLM, and a certificate-template review that removes ENROLLEE_SUPPLIES_SUBJECT where it is not needed. Require manager approval for higher-risk certificate enrollment.
  • Stop application database accounts writing files, changing server settings or running OS commands. One reported intrusion turned a database engine's own query logging into a web shell by pointing the log file at a web-served path — the privilege that allowed it had no application use.

Sources: FBI, 13 November 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026; Akamai, 23 October 2026 (via FBI FLASH note 1; link reaches their index)

Vendor trust

Why here

One compromised vendor's applications carried an actor into attempted sign-ins against more than 2,250 downstream tenants. Nothing had to be exploited at the customer: the applications were already approved.

The test

Name every third party holding a credential, OAuth grant, session token or federation trust in your estate, and the scope of each. Then ask how long it would take to answer 'what could they reach?' if one of them announced a breach this afternoon.

Passing looks like: The register exists and is current, and the answer is hours rather than a project.

What to do

  • Keep a register of every third party with credentials, grants, trust or data, with scopes and owners.
  • Review admin-consented applications, remove stale ones, and cut scopes to what the integration actually uses.
  • Assume vendor-held secrets are exposed the day a vendor reports an incident. Rotate then, rather than waiting for their final scope.
  • Watch for persistence after a vendor incident: credentials added to existing applications, new owners, role assignments, and applications assigning themselves privileged roles. These survive the original application being cleaned up.
  • Treat any agentic testing harness you run yourself as hostile code: isolated from production, no production credentials in it, scope enforced in configuration rather than in a prompt.
  • Ask your own suppliers which agentic offensive tooling they run against you or on your behalf, and where they obtained it. The original framework's repository was withdrawn in October, so the copies in circulation are forks and mirrors of unknown provenance — a supplier running one has introduced a supply-chain question, not just a testing practice.

Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 21 October 2026; Cloud Security Alliance, November 2026 (cited, not read); Sophos, November 2026 (cited, not read)

Money movement

Why here

The fastest route from access to money is a person, and the impersonation is now good enough that the control cannot be somebody noticing.

The test

Ask your finance lead what happens when a payee's bank details change on an existing invoice, and who can shorten that process.

Passing looks like: A callback to a number already on file, every time, with no seniority exception. If the answer depends on who asked, the control is one person's judgement.

What to do

  • Require out-of-band verification for any new or changed payee, bank detail change or urgent transfer — to a number already on file, never one supplied in the request.
  • Add dual approval and thresholds to payment workflows, with holds on first-time payees above a set amount.
  • Write down that a voice or video call from an executive does not authorise a transfer on its own, and agree a second channel for genuine urgency.
  • Know where sensitive data lives, alert on bulk reads and large outbound transfers, and line up counsel and incident response before you need them.

Our own recommendation. Not drawn from the reporting cited elsewhere on this page.

Detection and response

Why here

These agents are noisy and fast, which is an opportunity and a problem in one sentence. Static indicators help for days; behaviour holds after the infrastructure rotates.

The test

Find the last time a detection fired, was triaged, and was contained without a person being woken up. Then measure how long containment waits when the person has to be woken up.

Passing looks like: Containment for high-confidence signals is pre-approved and runs without a human in the path; the human approves anything irreversible. If every containment waits for a person, the response is slower than the attack by design.

What to do

  • Build the behavioural detections rather than relying on address lists, and tune each against your own baseline.
  • Inventory your own legitimate administrative capability before an incident rather than during one: which platforms can execute remotely, deploy, run scripts, issue certificates, synchronise identity or send to everyone. Those are the tools an intruder uses, because they are already approved, and you cannot restrict or monitor a capability you have not listed.
  • Keep the telemetry the story needs: interactive and non-interactive sign-ins, service-principal sign-ins, directory and cloud audit, API activity, object-storage data-plane, DNS and proxy.
  • Watch production egress for request-capture and out-of-band testing infrastructure, tunnelling providers, public file-transfer services and external model APIs. One reported run posted command output to 491 rotating capture URLs on a single service.
  • Pre-approve containment for high-confidence signals: disable the identity, revoke sessions and tokens, rotate the credential, block the source, isolate the host.

Sources: FBI, 13 November 2026; Microsoft Threat Intelligence, 21 October 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026

Testing

Why here

This class of tooling is being pointed at you whether or not you run it. The only question is whether you or somebody else sees the result first.

The test

Name the last time your external surface was tested end to end rather than scanned, and name one thing it found that you did not already know.

Passing looks like: Both answers exist, and the date is recent enough that the surface has not meaningfully changed since. 'A scanner runs weekly' is not an answer to this question.

What to do

  • Run authorised testing against your external surface now, and then on every significant release.
  • Point the testing at the overlooked systems — partner portals, mobile backends, employee apps, legacy APIs, anything built outside the core team.
  • Ship patches in hours rather than sprint cycles, and measure time from finding to fix in production.

Sources: FBI, 13 November 2026; Google Threat Intelligence Group and Mandiant, 30 October 2026

Scope

What this guide does not cover

Where this guide stops, and what it is not.

Subject

Covered

Not covered

Provenance

Covered

Public reporting from the FBI, Microsoft, Google and Mandiant, CrowdStrike, Zscaler, Censys and Akamai, cited per row.

Not covered

Anything of our own. RIA Labs has not observed this activity and has no telemetry on it. Nothing here is our finding.

Efficacy

Covered

Controls reported by the primary sources as relevant to this activity.

Not covered

Any ranking by measured effect. Nobody has published what proportion of these intrusions each control actually stopped, and we are not going to invent it.

Surfaces

Covered

Web applications, APIs, identity, and cloud control and data planes — which is where the reported activity landed.

Not covered

Operational technology and industrial control systems, and mobile or native binaries in any depth.

Indicators

Covered

Indicators we traced to a primary report we read ourselves.

Not covered

Indicators circulating in secondary coverage we could not trace. Several are in wider circulation and are deliberately absent.

Freshness

Covered

Checked against its sources on 18 November 2026.

Not covered

Any guarantee that the network indicators are still live. They decay in days, and the dates on the rows are the only claim being made about them.

The other half: what to look for

Behavioural detections with the thresholds and the false positives stated, five hunting queries with their logic written out, the indicator table with a source and a date on every row, and the corrections we are carrying.

Open the hunting pack

A manual penetration test was always a sample: two people, two weeks, against millions of lines. The industry lived with that partly because the attacker was sampling too — and that symmetry is what just ended. Running the same class of tooling against your own estate, with authorisation, is the last row of the matrix above. How ours works, or point it at your external surface.