Production secrets and cloud data were exposed in attacks linked to self-hosted AntiFox systems. The open-source AI penetration-testing platform was used from Alibaba Cloud infrastructure against internet-facing web services and AWS environments. Attackers deployed a webshell, examined application configuration, and used valid AWS identities for rapid discovery. Some activity succeeded in retrieving production secrets or reading, copying, and deleting S3 objects.
What Happened
Attackers used self-managed instances of AntiFox, an open-source agentic penetration-testing platform, in confirmed malicious activity against internet-facing web services and related AWS environments. The AntiFox infrastructure was hosted on Alibaba Cloud and supported tasks including vulnerability exploitation, webshell deployment, credential validation, configuration discovery, and AWS secret harvesting.
In one documented case, attackers placed an ASPX webshell on an internet-exposed IIS server. The IIS worker process then launched commands to inspect the host, list files, check permissions, and test whether files could be written into the website directory. The attackers used PowerShell to read the application’s web.config file and looked for database connection details and application settings. They also reviewed debug and error-log locations and copied configuration and log files into a web-accessible directory. The reporting did not identify the vulnerability used to gain the initial IIS access.
The AWS activity used valid IAM users and, in some cases, assumed EC2 roles. The attackers checked identities and permissions, then rapidly queried numerous AWS services. In one intrusion, they cycled through multiple identities and made hundreds of API requests in about 30 minutes.
Many requests were denied, but the attackers continued testing services and resources that remained accessible. Some sequences reached production secrets. In a separate incident, a valid AWS identity was used to list, read, copy, and delete S3 objects.
The evidence ties malicious activity to systems running AntiFox, but it does not establish how every valid AWS credential was obtained. In some cases, the starting point was a public IIS application; in others, the activity began with a valid AWS credential obtained through an unknown method.
Why It Matters
AntiFox illustrates how agentic security tools can make familiar attack techniques faster and more consistent. The individual actions—placing a webshell, reading configuration files, checking permissions, enumerating cloud services, and requesting secrets—are not new. The change is the ability to chain and repeat them with less manual effort.
The incidents also connect web security and cloud identity. An exposed application may contain configuration details or credentials, while a valid cloud identity can open a separate path into storage, secrets, databases, or other services. Defending those areas as unrelated systems can leave the transition between them difficult to see.
The large number of denied AWS requests is especially important. A high denial rate can look like failed reconnaissance, yet the observed attackers continued until they found permissions that worked. A small number of successful calls may be enough to retrieve a production secret or alter stored data.
For leadership, the practical issue is not whether the attacker used an AI label. It is whether access controls, telemetry, and response processes can contain rapid, automated discovery across multiple environments.
What to Watch
Defenders should watch for additional cases in which agentic penetration-testing frameworks are hosted in public cloud infrastructure and used against live systems. The supplied reporting already connects AntiFox to confirmed malicious actions, but the scale and range of operators using it remain unresolved.
Within affected environments, watch for AWS activity from hosting providers or source addresses that an identity has not used before. One address cycling through several keys, users, or roles in a short period is particularly relevant. Rapid List, Get, and Describe operations across unrelated services can indicate broad discovery even when most calls fail.
On IIS systems, watch for worker processes launching command shells, PowerShell, permission tools, or archiving utilities. New ASPX files in upload or temporary directories and configuration files copied into web-accessible locations also warrant investigation.
Teams should also watch for cloud activity beginning soon after suspicious web-server behavior, because that timing may reveal movement from an application compromise into AWS.
What Defenders Should Do
Web detection: Monitor for the IIS worker process starting unexpected command-line tools or PowerShell. Investigate new ASPX files in upload, temporary, image-processing, or conversion directories and any attempt to place configuration or log files in locations reachable through the website.
Cloud detection: Use AWS logging to identify unusual identity activity, especially rapid discovery across many services, access from unfamiliar hosting providers, or one source using multiple credentials. Review successful API calls within bursts of denials rather than closing the alert because most requests failed.
Secrets: Alert on unusual access to stored secrets and inspect whether application files expose database connection strings, cloud keys, or other credentials. If configuration material was accessed, assume the associated secrets require review even when there is no immediate evidence that each one was used.
Containment: If an IIS process launches an unexpected shell, isolate the server and preserve logs and forensic evidence. Remove malicious files only after evidence is retained and assess whether the activity reached internal databases or connected cloud services.
Identity response: Revoke compromised AWS sessions and access keys, rotate credentials exposed through application files, and review permissions for affected IAM users and EC2 roles. Apply least privilege so one valid identity cannot enumerate or access unnecessary services.
Data protection: Monitor access to sensitive S3 storage. Versioning and Object Lock can help protect important objects from deletion where those controls fit the organization’s operating requirements.
The Larger Lesson
AI-enabled attack tooling does not remove the importance of basic controls; it raises the cost of weak ones. An automated agent can test many permissions and services quickly, but it still depends on reachable applications, valid identities, exposed secrets, and excessive access.
The durable defensive model is therefore cross-environment visibility. Web telemetry, endpoint process activity, identity logs, cloud API calls, and secret access need to be connected closely enough to reveal the sequence rather than isolated alerts.
The incidents also challenge a common assumption that mostly denied requests indicate success for the defender. Denials are useful, but the attacker may need only one permitted action. Response teams should measure what succeeded, what data or secrets were reachable, and whether the same identity can continue elsewhere.
