Hackers Are Turning AI Agents Into Automated Credential-Hunting Machines
Amazon Detects AI Agents Using Stolen API Keys to Find More Credentials
Artificial intelligence is becoming a powerful tool for businesses. It is also becoming a powerful tool for cybercriminals. A new campaign described in records collected by Amazon shows how attackers can use AI agents to search for stolen credentials, test them and reuse working API keys to access more AI services.
Hackers Are Turning AI Agents Into Automated Credential-Hunting Machines
The important change is not simply that hackers are using AI. It is that AI agents can divide a complex operation into smaller tasks, share information and continue working with less direct human involvement. For businesses adopting AI automation, this creates a security problem that cannot be solved by protecting passwords alone.
Amazon's Sensors Capture an AI-Powered Credential Campaign
Amazon says records captured by its MadPot decoy sensors between October 3 and November 5 described agents using stolen AI API keys to power further credential-hunting activity.
There is an important qualification. Amazon did not observe the attacks directly and could not independently verify the outcomes described in those records. It assesses with moderate confidence that the agents used ARTEX, an open-source AI-powered penetration-testing framework.
The FBI separately warned in a November 13 advisory that some intrusions involving ARTEX had moved from initial targeting to privileged access in less than three hours.
These reports point to a faster and more automated attack process, although the available evidence does not establish that every ARTEX operation follows the same pattern.
ARTEX Turns Instructions Into Coordinated Actions
ARTEX translates natural-language instructions into commands performed by specialized AI agents. Rather than operating as a single chatbot, the framework can divide work among agents responsible for planning, execution and verification.
It relies on external large language models instead of supplying its own model. Amazon's records describe agents sharing information through a persistent knowledge graph, allowing information discovered during one stage to support later tasks.
This structure matters because attackers do not need to manually repeat every step. An agent can search for exposed services, inspect configurations, test credentials and record useful findings.
The technology is not inherently malicious. Automated penetration testing can help security teams identify weaknesses. The risk comes from applying similar capabilities to systems without authorization.
LiteLLM, Dify and Langflow Become Targets
According to Amazon's records, the campaign searched internet-exposed AI infrastructure and application-building tools, including LiteLLM, Dify and Langflow.
The reported targets included exposed management interfaces, unclaimed installations, default administrator credentials and provider API keys. The records also referenced the Langflow vulnerability CVE-2025-3248 and LiteLLM's master key.
The attackers reportedly tested discovered credentials and stored working keys in a shared pool.
This creates a dangerous cycle. A compromised AI gateway can expose credentials that provide access to other AI services. Those credentials can then support additional searches and operations.
For companies building AI applications, an API key is not just a technical setting. It can represent access to paid computing resources, private data and business-critical systems.
The Numbers Show a Wider Security Problem
The reports describe several related incidents and indicators of scale.
The FBI documented ARTEX-related intrusions in which attackers identified more than 110,000 objects in Active Directory and attempted to move toward domain controllers using stolen authentication material. Endpoint protection blocked at least one attempt.
The FBI also cited attacks on South Korean financial institutions involving the theft of personal information belonging to approximately 67,000 people.
Separately, Microsoft reported attempts to sign in to more than 2,250 downstream Microsoft 365 tenants using compromised ConnectWise Cloud Backup application credentials. Some attempts advanced to reconnaissance and credential changes. Microsoft cautioned that successful authentication did not necessarily mean customer data had been stolen.
These incidents should not be treated as one confirmed campaign with identical methods or outcomes. They demonstrate different ways compromised credentials and automation can threaten connected business systems.
AI Agents Change the Speed of an Attack
CrowdStrike's Adam Meyers described attackers issuing dozens of commands per hour, while the highest agent-driven activity observed reached approximately 1,500 commands in an hour.
That difference changes the economics of cybercrime.
Human attackers have limited time and attention. Automated agents can repeat searches, process results and execute multiple tasks at a much higher rate. A weakness that might once have required hours of manual investigation can become part of an automated workflow.
Meyers also described indications of agents reverse-engineering mobile applications to extract static keys. In another reported example, an attacker asked Claude where stolen information could be sold, including requests for Korean Telegram groups.
AI does not eliminate the need for human operators. It can, however, make their decisions and instructions more powerful.
The Business Model Behind AI Cybercrime
The underlying business model is straightforward. Attackers seek valuable access, automate the discovery of credentials and use compromised systems to expand their reach.
Stolen API keys can also create direct financial consequences. Unauthorized users may consume paid model capacity, generate unexpected API charges or access services connected to an organization's account.
For legitimate businesses, this means AI spending and AI security are becoming connected management issues. An unexplained increase in token consumption may be more than a billing problem. It could indicate that a credential has been exposed.
The source material does not provide a confirmed financial estimate for the campaign or the total value of stolen credentials. That information remains an important gap.
The Open-Source Problem Is Also a Governance Problem
The FBI says ARTEX's developer removed the original GitHub repository in October after reports of malicious use. However, forks and archived copies remained available. Researchers identified more than 1,600 publicly exposed ARTEX consoles by early November, according to the FBI.
Removing an original repository does not remove software that has already been copied. Open-source distribution makes useful security tools accessible, but it also makes it difficult to control how those tools are reused.
The central policy question is not whether AI penetration-testing frameworks should exist. Security researchers and defenders need automation too.
The harder questions concern accountability, authorization and safeguards. Who is responsible when an agent performs an action outside its intended scope? Can its commands be audited? Can organizations identify the person or process that authorized access? Can dangerous operations be stopped before they spread?
A tool's open-source status does not make unauthorized access acceptable. At the same time, restricting software alone cannot replace sound security controls.
Small Businesses Need to Treat AI Credentials as Critical Assets
Smaller companies may assume that sophisticated AI-powered attacks mainly target banks, cloud providers and large technology firms. That assumption is risky.
A small business may expose an AI gateway, leave a default administrator password unchanged or store API keys in application code. These weaknesses can provide attackers with access to paid services and connected systems.
The FBI recommends inventorying internet-facing applications, checking for exposed secrets and monitoring changes to credentials and permissions. Organizations should also monitor unfamiliar LLM proxies, unusual token usage, unexpected API charges and suspicious connections to model providers.
For smaller teams, practical priorities include:
- Keep AI gateways and management interfaces off the public internet unless access is necessary and properly protected.
- Replace default credentials and rotate API keys that may have been exposed.
- Store secrets in controlled secret-management systems rather than source code or public configuration files.
- Apply least-privilege permissions so each application receives only the access it needs.
- Set spending limits and alerts for unusual API consumption.
- Keep logs of credential changes, agent activity and access to sensitive systems.
- Review third-party integrations and remove credentials that are no longer required.
The FBI also cautions against blocking IP addresses solely because they appear in its advisory. Cloud addresses can be reassigned, so security teams should investigate the wider activity rather than treating an IP address as proof of malicious behavior.
The Weakest Point Is Often the Connection Between Systems
The biggest lesson from the ARTEX reports is that AI security cannot stop at the model itself.
A business may choose a trusted AI provider and still be exposed through an unsecured gateway, a compromised integration or an administrator credential that has been left unchanged. The weakest point may sit between systems rather than inside the AI model.
As companies connect AI agents to databases, cloud platforms, customer records and business applications, they must treat every connection as a potential security boundary.
AI agents can make businesses faster. They can also make mistakes and misuse credentials at machine speed. The organizations that benefit most from automation will be those that combine useful autonomy with clear permissions, monitoring and human accountability.
The future of AI security will not be decided only by which model is most intelligent. It will depend on who controls what the model can access, what actions it can take and how quickly suspicious behavior can be stopped.
Hackers Are Turning AI Agents Into Automated Credential-Hunting Machines
AI cybersecurity, ARTEX framework, stolen API keys, AI agents, credential theft, Amazon MadPot, LiteLLM security, Dify security, Langflow vulnerability, FBI cyber warning, AI security risks, API key protection, cloud security, automated cyberattacks
Share: