The attacker is an agent now: what ARTEX means for cloud identity
AI-powered attackers are changing the cloud threat landscape. ARTEX, an open-source agentic framework, demonstrates how attackers can automate planning, credential discovery, and exploitation using cheap LLM APIs, gaining privileged access in under three hours. Discover how traditional attack techniques have evolved, why detection alone is no longer enough, and eight practical steps to harden your cloud environment.
The attacker in your cloud account is most-likely no longer a person. In one documented case, an automated agent was reported by Microsoft reading Key Vault secrets about five minutes after a GitHub secret search, using nothing but a valid credential.
That is the lesson of ARTEX, an open-source agentic framework that the FBI, Microsoft, Amazon and Google all reported on this autumn, with several campaigns targeting cloud environments. Nothing in its attack chain is a new technique. What changed is speed and cost: planning, key hunting and exploitation now run at machine pace on cheap LLM APIs, and the FBI and Microsoft report privileged access in under three hours in some cases.
Detection alone is too slow against an attacker that operates as an agent with valid credentials. It needs prevention behind it: controls at the identity and network layers.
Here is how a pentest tool became a cloud weapon in four months, why nothing about it is new, and the eight steps at the end that harden your cloud environment against this kind of attack.
From pentest tool to cloud weapon in four months
ARTEX is an open-source Go framework that turns a plain-language goal into commands run by a team of AI agents with little human input. It has no model of its own and calls outside LLM providers, most often low-cost Chinese-origin models, according to the FBI and Check Point, although the FBI lists Claude and GPT among supported providers. It was released on GitHub in July 2026 for authorized testing. Abuse began within weeks, and November is when it became a mainstream cloud problem: the developer withdrew the repository on Oct 8 and went closed source, but forks and a commercial version remain, and about 1,600 ARTEX consoles were live on the internet in early November.
| Date (2026) | What happened |
|---|---|
| July | ARTEX released on GitHub |
| August | FBI cites malicious use from at least August; Microsoft dates the mass GitHub secret searches to August |
| Late Aug to Oct | Agents test stolen AWS and Azure keys; Mandiant sees ARTEX and similar frameworks in incident response from late August |
| Oct 8 | Developer withdraws the original repo and goes closed source; 12,000+ forks already exist |
| Oct 10 to 16 | Password spray against 327 accounts at eight organizations (Oct 10 to 14, no successful sign-in); ConnectWise Cloud Backup compromise reported publicly Oct 15, advisory Oct 16 |
| Oct 21 | Microsoft reports Storm-0940 using ARTEX to attempt sign-ins to 2,250+ tenants through a compromised ConnectWise app; Wiz (Oct 22) and Google (Oct 30) publish overlapping reports |
| Oct 28 to Nov 3 | Leaked Entra app secrets used for service principal sign-ins (Microsoft’s earliest sighting Oct 28, Todyl’s Nov 1); in one undated Microsoft case, a Key Vault was read about five minutes after a GitHub search |
| Nov 5 to 6 | Microsoft publishes its exposed-secrets report and Check Point its console research (1,608 live of 2,557 hosts seen); Todyl updates its Entra investigation on Nov 6 |
| Nov 13 | FBI FLASH advisory and Amazon Threat Intelligence report published |
What made it dangerous to cloud specifically was the combination of an automated attacker and long-lived credentials. Cloud APIs accept a valid key from anywhere unless a network or identity control says otherwise.
What the campaigns do
The public reporting covers several separate campaigns, including Storm-0940’s ConnectWise supply-chain compromise; activity Google tracks as UNC7228, which overlaps it; Storm-3358, which the FBI cites for WISP command-and-control infrastructure; unnamed actors abusing leaked Entra secrets; a financially motivated attack on South Korean financial institutions that exposed about 67,000 people’s data; and Check Point’s cybercrime cases targeting Apple signing certificates, crypto payment platforms and gaming credentials.
Microsoft, Google, Amazon, Check Point and Todyl describe the same pattern across these campaigns. The steps below come from different actors and are combined here to show it. Every step is a legitimate API call made with a stolen credential.
- Find the secret. Agents run mass GitHub searches pairing organization names with secret keywords, and harvest provider keys from exposed AI gateways such as LiteLLM and One API. Amazon notes keys exposed for more than 30 days are treated as low-value, so speed matters on both sides.
- Sign in as a workload. Leaked Entra client secrets are used for service principal sign-ins from rented hosting-provider servers (VPS). No user and no password, and MFA does not apply to this flow by design. AWS IAM user keys are tested with DryRun calls, 27 events in 23 seconds of rapid permission probing in one Google-reported operation.
- Turn one secret into many. Azure Resource Manager operations that return credentials expose storage keys, App Service publishing credentials and connection strings. Agents then read Key Vault, run commands through Kudu and VM Run Command, and use Run Command or added app credentials for persistence. Microsoft also saw agents searching Azure DevOps for secrets.
- Fan out through vendors. A stolen secret for a multi-tenant vendor app yields tokens in every tenant that consented to it. In the ConnectWise case, Microsoft reports Storm-0940 attempting sign-ins to 2,250+ downstream tenants, and an app self-assigned Global Administrator. Todyl reports one MSP app sweeping 15 client tenants in about 10-second loops.
- Take the value. On AWS, Google reports agents building jump hosts and staging EBS snapshots, and Check Point reports a separate cybercrime campaign overwriting a public S3 bucket with leaked credentials. Amazon assesses that stolen model keys were likely pooled so victims pay for the attacker’s inference. Todyl counted 81 SharePoint files and about 2,279 mailbox items accessed.
Amazon says it could not verify agent-reported outcomes, and neither Amazon nor Google confirms that staged AWS data left the account.
Known techniques, machine speed: why prevention holds up
Every step in the reported campaigns is a technique defenders already know: a leaked secret, a valid sign-in, a credential-returning API call, a Key Vault read. ARTEX adds speed and scale, not new capability. That is what makes the outcome controllable. A control that refuses a stolen secret from an untrusted network, removes the public path to a secret store, or disables long-lived keys works the same against an agent as against a person.
What changes is how little time that leaves a defender. Three facts from the reporting show why detection alone loses to an agent. First, speed: the FBI and Microsoft cite under three hours from initial targeting to privileged access, and Google describes execution bursts of 2 to 25 seconds. Second, there is nothing to detect at the endpoint. The activity is valid sign-ins and API calls, and Todyl notes that detections keyed only on successful sign-ins miss the failed attempts that precede them. Third, agents adapt: Todyl saw the operator switch from curl and Python to spoofed browser user agents once detections fired, and Wiz saw new IPs appear within minutes of blocks.
By contrast, an environment with the right preventive guardrails in place is a poor target for ARTEX, however capable the attacker’s tooling becomes. A guardrail is enforced before the request succeeds, so the attacker’s speed stops mattering. This is the lesson AI-assisted attacks keep teaching: when attackers get cheaper and faster, the defense that holds is prevention already in place, not detection that has to win a race. Two cases need extra care: consented vendor apps, and attackers operating from inside allowed network ranges.
How much any control helps depends on where the attacker operates from. Workload-identity Conditional Access and the Key Vault firewall can refuse a request outright when the credential is used from outside allowed networks, while most other controls only limit what a stolen credential can reach. Vendor multi-tenant apps call for a different control: a consented app already holds the rights it was granted, so review its permissions and remove consent you no longer need. Prevention here means containment: shrinking where a stolen credential works, not making it harmless.
So the question to ask is not how fast you can respond. It is which non-human identities can sign in from where, and which secret stores they can reach.
What breaks the chain: ARTEX mapped to Blast guardrails
Blast’s threat-intel catalog maps each campaign to deployable guardrails, and it now holds more than 100 records. For ARTEX we selected 19 guardrails. Most restrict where or how a stolen credential can be used, and none makes a stolen credential harmless. Each was kept only where its mechanism limits a documented step, not just where a MITRE label matches; the metadata row is hardening for adjacent techniques. Where a control has a known limit, the row says so.
| Attacker action | Blast guardrail | Why it matters |
|---|---|---|
| Stolen service principal secret used from a VPS | Denies workload identity sign-ins from untrusted locations or flagged as high risk. | A leaked secret used from a VPS is refused, because the sign-in comes from outside your allowed locations. Strongest for single-tenant workload identities. |
| Password spray and replayed user tokens | Requires MFA and limits user sign-ins to trusted locations. | A sprayed password fails without the second factor, and replayed tokens are refused from outside your locations. |
| Key Vault reads | Keeps Key Vaults off the public internet by requiring a firewall. | Agents can no longer read secrets over the internet, because the vault accepts traffic only from approved networks (trusted Azure services excepted). |
| Storage keys and data | Disables account-key access to Storage and limits it to approved networks. | Harvested storage account keys stop working, and the data stays reachable only from approved networks. |
| Stolen AI model access | Requires identity-based access to AI services and keeps model calls inside trusted networks. | A stolen model key is useless from outside, because access needs a valid identity and calls must come from trusted networks. |
| AWS stolen IAM keys | Keeps AWS identities and role assumptions inside trusted networks and rules out long-lived keys without MFA. | Stolen IAM keys fail from outside trusted networks, and long-term keys cannot be used at all without MFA. |
| Secrets and S3 | Allows secret and bucket access only from trusted networks. | Stolen credentials cannot read secrets or overwrite buckets from the attacker’s own infrastructure. |
| Metadata and footholds | Hardens instance metadata access and limits the footholds an attacker can set up. | Credential theft through instance metadata is blocked, and attacker jump hosts become harder to set up. |
Network controls fail if the attacker operates from inside an allowed range or replays a token from a trusted host, so roll them out in audit mode first and keep break-glass exceptions.
Immediate steps to harden cloud identity, secrets and non-human identities against agentic attacks
- Scan your public and private repos, images, pipelines and CI logs for client secrets and rotate any found, and turn on GitHub push protection so supported secrets are blocked at push. Deleting a secret from a repo does not revoke it.
- Put single-tenant workload identities under Conditional Access with named locations for their egress ranges (Workload Identities Premium; start in report-only mode).
- Block new client secrets with app management policies, rotate or remove existing ones, alert on credentials added to existing apps, and move workloads to managed identities.
- Put Key Vault and Storage behind firewalls or private endpoints, and turn off Storage shared key access.
- Turn off basic authentication for App Service deployment (SCM and FTP) and restrict network access to the Kudu site.
- Require IMDSv2 and a hop limit of 1 on EC2.
- Inventory consented multi-tenant vendor apps in your tenants, review their permissions, restrict user consent, and flag apps that hold RoleManagement.ReadWrite.Directory, and limit app access to SharePoint sites and Exchange mailboxes with Application Restrictions and RBAC for applications.
- Watch AI-service spend and set budget alerts and quotas, since stolen model keys are billed to you. On AWS, also replace IAM users with roles where possible and enforce a maximum access key age.
The next ARTEX will not announce itself, and it will not need a new technique. The environments that stay safe will be the ones where prevention was already in place. Make a stolen credential as useless as you can. If you want help with any of this, whether it is a question about your own environment or putting these controls in place, reach out to us. We are glad to help, and you can book a demo to see how Blast maps this campaign and the rest of our catalog to guardrails you can deploy.