ARTEX AI provider key campaign is the clearest finding in Amazon Threat Intelligence’s 13 November 2026 report. Its MadPot decoys received AI-agent requests from 3 October through 5 November that carried campaign memory. Those records describe a loop: find provider keys, test their access, and use the keys to fund further model calls. Amazon observed the requests to its decoys; it could not verify the agents’ reported outcomes at other organizations.
This is a useful case for teams that operate AI gateways or application builders. The exposure sits where an internet-facing application can reach valuable upstream provider credentials.
A cyber threat intelligence platform can help analysts relate a reported campaign to their own observed activity. The investigation question is whether a local gateway, key store, or billing account shows the same sequence. Our ARTEX framework profile explains the tool; this report examines Amazon’s newly described credential-hunting campaign.
What Amazon Actually Observed
Amazon configured MadPot sensors to resemble common LLM inference interfaces. ARTEX-linked agents sent prompts to those sensors, including task context, tool histories, and reported findings. Amazon assesses with moderate confidence that ARTEX powered the campaigns, based on references in prompts and related technical artifacts. The records reveal instructions and agent memory, but they are not independent telemetry from every target the agents described.
Amazon says it assessed activity referring to potentially affected AWS customers and notified them where appropriate. It explicitly says activity against other organizations could not be confirmed. The distinction matters when deciding whether a named product or organization in agent memory suffered a breach. A reported success inside an agent transcript calls for local verification before it becomes an incident finding.
The same report describes a separate attempt to bypass identity checks at cryptocurrency exchanges. That effort and the provider-key campaign shared an agent framework, but they had different goals. Amazon says the recorded deepfake attempts were rejected by the verification service. The key-hunting operation is the focus of the checks below.
How the ARTEX AI Provider Key Campaign Worked
The operators gave the agents standing instructions, including target priorities and a way to grade found credentials. Amazon describes planner, worker, verification, and compression agents sharing a persistent knowledge graph. That shared memory let one agent continue from another’s reported result. It also preserved the history that later appeared in requests to MadPot.
The agents searched for exposed AI applications and material left in public code or packaged artifacts. They then sought the upstream provider keys held by gateways and application builders. The records describe checking whether a key worked, what permissions it had, and whether it could reach premium models. Amazon says stolen keys were placed into the framework’s own model gateway to power more searches.
- DiscoverFind exposed AI applications and public artifacts.
- Reach a key holderLook for upstream provider credentials in gateways or builders.
- Validate accessCheck whether a credential works and what it can reach.
- Reuse capacityFeed usable keys into the agents’ model gateway for more searches.
This flow summarizes what Amazon found in agent records. It does not establish that every described key was successfully stolen or charged. For comparison, our earlier ARTEX exposed-secrets analysis covered a different source and attack angle. Amazon’s MadPot material adds the reported self-funding loop and its direct visibility limit.
Why AI Gateways Concentrate the Risk
Amazon names gateways and builders, including LiteLLM, Dify, Langflow, and RAGFlow, as product classes in the agents’ targeting records. A gateway can hold credentials for several upstream models. One exposed management path or secret store may therefore affect more than one provider account. Naming these products describes the agents’ interests; it does not mean every installation was compromised.
The records list several routes to an exposed secret: default or hardcoded credentials, unauthenticated management access, unpatched software, and credentials published in code or artifacts. Amazon also describes cloud credentials as a path to managed AI services and secret stores. These are different entry conditions, so a team should inventory its actual deployments before selecting a fix. The root issue is privileged provider access sitting behind a reachable, weakly controlled interface or artifact.
From 18 August to 16 November 2026, Brandefense Blackhole counted 268 dark-chat posts matching ARTEX. The unit is matching posts, not actors, incidents, or successful thefts. The count covers ARTEX discussion across that period and cannot be assigned to Amazon’s specific campaign. It is context for monitoring interest, while local access and billing records decide whether a key was misused.
The source report alone cannot identify which local accounts require rotation. Teams need gateway configuration, identity, provider, and billing evidence to make that decision.
What Should Defenders Check First?
Start with an inventory of internet-reachable gateways, AI builders, inference interfaces, and the provider accounts each can use. Identify where keys are stored, who can read or export them, and whether administrative access still uses defaults. Compare that inventory with the conditions Amazon describes. Prioritize a single interface that holds several production keys over an isolated test account.
Next, review provider audit and billing records for unfamiliar model use, unusual spend, or access from an unexpected application identity. Check gateway administration and secret-store reads in the same period. A spike alone is not proof of this campaign; correlate it with deployment changes and authorized workload. Preserve the relevant records before rotating a suspected credential.
- Map key holdersRecord each gateway or builder and the upstream provider accounts it can reach.
- Close obvious exposuresRemove default credentials, restrict management access, and update affected software.
- Compare usageReview provider calls, spend, gateway administration, and secret-store reads together.
- Rotate suspected keysRevoke exposed credentials and replace them in the applications that need them.
- Verify containmentConfirm old keys fail and expected workloads use only the replacement keys.
These are our defensive steps drawn from Amazon’s described exposure paths and advice to watch unusual AI-service usage. Rotation should include a check that old credentials can no longer pay for inference. If a provider key had broader cloud permissions, review its role and accessed services before closing the case. A guide to AI agent identity and access adds context for restricting non-human identities.
Where the Evidence Stops
Amazon’s strongest evidence is the content of requests that reached its decoys between 3 October and 5 November 2026. The reported actions before and between those requests occurred elsewhere. Amazon could read the agents’ accounts of them, but could not watch or confirm every claimed outcome. Its moderate-confidence ARTEX attribution is an assessment, not a verified identity for each operator.
Defenders can act on the exposure pattern without treating an agent’s success claim as a confirmed breach. Check actual key access and spend, then record which observations support any incident decision. This keeps a useful campaign warning separate from the narrower question of what happened in one environment.
Frequently Asked Questions
Did Amazon confirm every key theft described by the ARTEX agents?
No. Amazon observed requests to its MadPot decoys and read the agents’ records. It says outcomes against other organizations could not be confirmed.
Why were AI gateways attractive to the agents?
Amazon says gateways can aggregate upstream provider keys. One weakly controlled interface can therefore expose credentials for multiple downstream accounts.
Which records should a team review for suspected provider-key misuse?
Review gateway administration, secret access, provider API usage, application identities, and billing together. Correlate anomalies with authorized deployments before deciding whether a key was misused.
