Internet-facing AI applications can expose more than a chatbot. When an application stores credentials for multiple model providers, a compromised administrator interface can become a route into several paid AI accounts. A November 16 report from The Hacker News describes Amazon threat intelligence findings in which attackers’ AI agents sought out these credentials—and reported using stolen keys to support further attacks.
For teams running LiteLLM, Dify, One API or similar platforms, the practical response is familiar: restrict administrative access, remove default credentials, patch exposed services and monitor provider usage. The important change is the speed and persistence with which attackers can repeat those checks.
What Amazon observed—and what remains unverified
According to the report, Amazon’s MadPot decoy infrastructure posed as popular large language model services. Agents contacting those decoys submitted requests containing operational histories, including instructions and accounts of their findings. These records described searches for exposed AI applications, extraction and validation of provider keys, and aggregation of credentials into a shared pool.
There is a significant evidence boundary: Amazon did not directly observe the activity described on those other systems and could not verify the agents’ claimed outcomes. The records reveal attacker objectives and methods, not a confirmed compromise of every named target.
Amazon assessed with moderate confidence that the agents used ARTEX, an AI penetration-testing tool. Its assessment that separate China-linked groups operated the campaigns also carried moderate confidence. Neither assessment should be treated as certainty.
The named targets extended beyond LiteLLM and Dify to gateways, application builders and inference tools, including One API, New API, Langflow, RAGFlow and Open WebUI. Being named in an attacker playbook does not mean every deployment of a product is vulnerable.
Why provider keys are valuable targets
A model gateway often centralizes credentials so applications can access several providers through one service. That simplifies operations, but concentrates risk: administrative access may expose credentials with a much wider reach than the gateway itself.
The captured records described attackers prioritizing premium AI models and searching repositories, package registries, container layers and other public sources for secrets. They also described using stolen credentials to power the models directing further searches. Amazon could not independently confirm those reported outcomes, but the operational implication is clear: an AI provider key deserves the same protection as other production credentials.
Potential consequences include unauthorized inference charges and loss of service when quotas are exhausted. Broader exposure depends on the permissions attached to each credential; an inference-only key should not be assumed to grant access to unrelated cloud resources.
Close the easiest entry points first
Start with an inventory of externally reachable AI services. Include experimental deployments, abandoned proofs of concept and management interfaces—not just applications listed in your production catalog.
Prioritize these checks:
- Restrict administration. Put management interfaces behind private access or a strongly authenticated access layer. A public chatbot does not require a public administration console.
- Replace defaults and finish provisioning. Change default administrator credentials and master keys. Keep incomplete installations inaccessible until ownership and authentication are configured.
- Review unauthenticated routes. Check configuration, provider-settings, debugging and inference endpoints. Test access from outside your trusted network rather than assuming the frontend login protects every backend route.
- Patch against vendor guidance. The report identifies Langflow CVE-2025-3248 and RAGFlow CVE-2026-28797. Confirm affected versions and current remediation in the relevant vendor advisories; do not assume one upgrade recommendation applies to every product or fork.
- Constrain outbound requests. URL-fetching features can create server-side request forgery exposure. Review destination validation, egress controls and protections around cloud metadata services.
These are defensive recommendations, not evidence that your installation has been compromised.
Reduce the impact of a stolen key
Separate credentials by application and environment. Avoid sharing one powerful provider key across production, development and demonstrations. Use the narrowest permissions available and store secrets outside client-side code and public build artifacts.
Scan repositories, deployment configuration, CI/CD systems and container images for exposed credentials. If a key has been published, deleting the visible copy is not enough: revoke or rotate it, update legitimate consumers and investigate usage during the exposure window.
Where providers support them, configure spending alerts, usage limits and appropriate credential restrictions. Check whether a budget setting actually blocks consumption or merely sends a notification. Alerts are useful, but they are not automatically a financial containment mechanism.
Watch usage and automation accounts
Investigate unexpected inference spending, unfamiliar model use, unusual request volumes and activity from infrastructure your organization does not use. Correlate provider activity with gateway logs and administrative changes before deciding whether traffic is legitimate.
The Hacker News also describes a separate Cloudflare investigation involving ARTEX and a compromised CI/CD API token. Responders observed 400 database dump requests in an hour, but the report does not establish that those requests returned data. This illustrates why attempted access and confirmed data loss must remain separate findings.
Review automation tokens for excessive scope, missing expiration and absent source restrictions where supported. Investigate unexpected Worker deployments or other newly created cloud resources, but do not rely on tool-specific names alone: attackers can rename their artifacts.
If suspicious activity appears, preserve logs, restrict the exposed service and revoke affected credentials. Restore access only after addressing the entry point. Rotating a key while leaving the same administrative weakness accessible invites a repeat incident.
The priority is not buying a special defense against AI agents. It is making sure automated attackers cannot repeatedly exploit basic weaknesses in systems that hold valuable credentials.
Source: The Hacker News report on Amazon’s AI-agent campaign findings.