
The FBI ARTEX Warning Is a Call to Enforce Cloud Security Architecture
The FBI’s November 13 warning on ARTEX should put enforced cloud security architecture at the top of the AI defense agenda.
Microsoft and AWS have already documented ARTEX activity. Microsoft observed an agent go from searching GitHub for exposed secrets to extracting Azure Key Vault data in just five minutes. AWS’s MadPot honeypots captured agents’ campaign records describing credential theft and reuse. Microsoft disclosure · AWS disclosure
The FBI describes agentic attacks spanning reconnaissance, initial access, persistence, privilege escalation, data collection, and command and control. At Native, we analyzed 50 reported actions across eight phases and mapped them to native cloud and Kubernetes controls. We identified ways to block critical steps and contain their impact by restricting permissions, trust relationships, and network paths.
The priority is getting those controls in place before the attacker acts. FBI advisory
What is ARTEX?
Originally built for authorized security testing, ARTEX can use different large language models (LLMs). An operator sets an objective; the tool searches for weaknesses, tests possible entry points, evaluates results, and determines what to try next, with little further human input.
Public reporting describes its use by individual attackers, organized groups, and nation-state actors, from opportunistic scanning and credential theft to targeted intrusions and supply chain activity. Reported targets span financial services, cryptocurrency, government, telecommunications, defense, pharmaceuticals, semiconductors, and technology organizations across multiple regions. FBI advisory
ARTEX can probe defenses systematically, test paths in parallel, and adapt when blocked, finding weaknesses a human might miss or lack time to pursue.
We can no longer count on a gap going unnoticed or on having time to respond. The boundaries need to exist before the attack.
Cloud is one of the main targets
The FBI documents stolen cloud credentials being tested, cross-account trust abused, deployment interfaces used for execution, cloud metadata queried, and storage read in bulk. Attackers also used compute and volumes to stage collection and exfiltration. FBI advisory
A leaked credential or application weakness can open the cloud control plane, giving an attacker access to the capabilities legitimate teams use to create infrastructure, assign authority, and access data.
AWS’s CJ Moses described how MadPot honeypots posing as model inference APIs received agents’ operational histories. They showed credential hunting, tests of stolen credentials, and stolen model access used to power further hunting. AWS could not independently verify outcomes outside its own monitoring. AWS disclosure
The architectural question is simple: how much can an attacker do with the access they obtain?
A CSPM finding cannot deny an API call
Cloud security posture management (CSPM) can identify exposure, prioritize risk, and initiate remediation, including automated workflows. That matters. AWS CSPM documentation
But a finding cannot reject a request, revoke a permission, or close a network path. When an attacker uses a valid credential, only an effective restriction can stop the privileged action. If the response is another finding awaiting action, the attack can continue.
We should measure security by what the environment will actually prevent or contain, not only what it detects.
Turn attack techniques into enforced boundaries
The FBI’s reporting points to four practical priorities.
Separate public applications from administrative access
An application may need to be public; its deployment interface, database administration tools, and cloud management paths do not. Restrict those paths to approved identities and networks, and prevent new deployments from exposing them. The FBI explicitly recommends protecting management and publishing interfaces with trusted access paths. FBI advisory
Limit stolen credentials and third-party access
Grant workloads and integrations only the resources and actions they need. Restrict cross-account trust, remove unnecessary authentication methods, and govern who can create credentials or change permissions.
Revoke or rotate already exposed credentials: blocking new secrets does not invalidate existing ones, as Microsoft notes. Microsoft disclosure
One compromised identity should not provide authority over unrelated systems.
Keep workload access from becoming infrastructure administration
Reported techniques include launching compute with privileged IAM roles. AWS requires separate permission to pass a role to a service; restricting compute creation, role passing, and identity changes to approved operational identities can block this escalation path. Mandiant analysis · AWS role permissions
The principle applies across clouds: an application identity should not be able to create more privileged infrastructure or weaken its own boundaries.
Constrain data access and outbound communication
The FBI describes bulk storage reads, output sent to external endpoints, and command-and-control tunneling. FBI advisory
Scope each workload’s data access and enforce which destinations, protocols, and network paths it can use. A public-facing application does not need unrestricted outbound access.
No single control is sufficient. A compromised workload may still abuse data it legitimately reads or destinations it is allowed to reach. Layered cloud boundaries limit its options and complement application, identity, and endpoint defenses.
The hard part is making this true across complex cloud environments
The controls exist. Operationalizing them safely and consistently across clouds and clusters is the hard part.
Consider preventing ordinary workload identities from creating infrastructure or changing security controls. Teams first need to identify which applications and deployment pipelines genuinely require those permissions and which identities accumulated excess authority.
Then they must translate one security requirement into each cloud’s different policy language, identity model, and enforcement scope, making gaps explicit. They need to assess impact, test legitimate workflows, deploy changes in stages, and roll back safely when necessary. Otherwise, every new guardrail becomes a bespoke project.
Enforcement also needs continuous verification. Detect drift, uncovered resources, and changes that reopen access, and adapt the policy when services or deployment patterns evolve.
The goal is to enforce security architecture quickly, consistently, and at scale, and keep it effective as production changes.
The defenders advantage
I’m optimistic because we know what to do.
The speed and scale of AI-powered attacks are daunting. But our advantage as defenders remains: we define the environment attackers operate in and what they’re allowed to do.
It comes back to fundamentals: limiting exposure, segmentation, least privilege, and secure baselines.
The challenge is getting them right everywhere, all the time, across every cloud, workload, and change.
That’s what we built Native for: to turn security architecture into continuous, consistent enforcement at scale.
If you’d like an assessment of whether your environment currently has the controls in place to prevent and contain ARTEX-powered attacks, reach out to me directly at amit@native.security or request an assessment here.
Related content
More articles
Security Architecture & StrategyTurning Attack Research Into Controls: What We Learned Taking the Zealot Cloud Hack Apart
DateNovember 10 2026Read7 min
Native Team
Security Architecture & StrategyWhat We Learned Taking the Hugging Face Agent Breach Apart
DateOctober 20 2026Read8 min
Jeannie Christensen
Security Architecture & StrategyFrom Security Intent to Enforced Invariants
DateOctober 5 2026Read5 min
Ariel Septon
