Amazon says it observed agent-assisted cybercrime campaigns through requests sent to its MadPot decoy sensors. In a November 13, 2026 AWS Security Blog report, Amazon Threat Intelligence said that between October 3 and November 5 those requests included stored records from operations run with ARTEX, an open-source agentic penetration-testing framework. The report describes activity the sensors could see and records the agents sent; it is not an independent confirmation of every claimed intrusion or victim impact.
What did Amazon's sensors actually observe?
MadPot presents controlled services to potential attackers and records how they interact with them. In this case, Amazon says some agent requests reached decoy endpoints configured to resemble popular large-language-model inference services. The requests carried task history and campaign notes that the agents ordinarily used to coordinate work. Amazon could therefore read what the operators and agents said they were trying to do, even though it was not watching every target system directly.
Amazon's report describes two broad campaigns: efforts to steal credentials that could pay for further model use, and attempts to defeat identity checks at cryptocurrency exchanges. The company says AWS customers it assessed as potentially affected were notified. It explicitly says it could not confirm activity against other organizations and could not verify the outcomes the agents reported about systems outside MadPot's sensors. The campaign records show intent and claimed progress, not a confirmed count of successful victims.
How much can the report establish about the attackers?
Amazon attributes the operations to ARTEX with moderate confidence, based on references to the framework and related technical artifacts. It also assesses with moderate confidence that different China-linked threat groups were involved. Those are Amazon Threat Intelligence's assessments, not independently established identities. The company says the campaign records contained reports of activity across hundreds of targets, but it could not validate all of those claims.
The report's most important boundary is how the evidence was collected. The honeypot saw the agents when they contacted decoy model endpoints; the agents' persistent notes also described activity before and between those requests. Amazon says it did not observe those other actions as they happened and could not verify the agents' reported results. This makes the report useful for understanding how operators are trying to organize work, while leaving uncertainty about the success and full reach of those operations.
What should app and cloud operators take from it?
The defensive point is credential and access hygiene, not a new attack recipe. Amazon's report emphasizes familiar weaknesses such as exposed or hard-coded keys, insecure management interfaces and unpatched systems. Its advice is to treat AI applications, model gateways and inference endpoints like other production services, monitor unexpected usage or spend, and remediate exposed secrets quickly. AWS's IAM guidance also recommends temporary credentials for workloads, least-privilege permissions and regular review of unused keys and access.
For a maker shipping an AI feature, that means asking where provider keys live, which services can use them, whether permissions are narrower than the whole account, and how an unusual increase in model calls would be noticed. A credential that appears in a public repository or a misconfigured app can create both access and billing risk. Rotating exposed keys and checking access logs are sensible first steps, but teams should follow the current guidance from each provider and their own incident-response process.
Do AI agents change what defenders need to do?
Amazon argues that agents can make established techniques more persistent by dividing work and retaining shared task context. The report does not show that the underlying weaknesses are new, or that agents make every campaign successful. Its practical signal is that defenders should assume exposed credentials and weakly protected services can be found quickly, and that model usage itself can be part of the cost an attacker tries to shift onto someone else.
For software makers, the lesson is to build security into the ordinary app lifecycle rather than treating AI features as a separate exception. Protect credentials, limit what services can do, keep public management routes behind authentication, and have a way to detect abnormal usage. VIBPS is a social network for software where makers share apps and their ongoing work with an audience; protecting the infrastructure behind those apps is part of earning that audience's trust.
