An ARTEX operator sent AI agents after iOS app-signing certificates at ten gray-market signing shops. Over roughly two weeks, the agents obtained more than 300 certificates and passwords; three passed the checks needed for use in signing apps. The operator’s conversation with the agents mostly consisted of progress checks. Check Point Research’s account comes from exposed ARTEX operator databases, rather than the web-server logs defenders usually see after a break-in.
For teams that run app-signing services, credential stores or public web applications, this is a concrete example of an agent fleet finding entry points and pursuing a business goal after a person sets it. A stolen enterprise-signing key and its matching provisioning profile could let someone sign and distribute an iOS app outside the App Store. That makes the certificate store and the public APIs around it worth inspecting together.
A signing shop became the target
ARTEX is an AI-driven penetration-testing platform whose agents coordinate reconnaissance and security tools. Check Point found ten ARTEX installations exposing their full databases, including prompts, model configurations and attack records. Its wider survey identified 2,557 systems that hosted ARTEX panels at some point, including 1,608 live, publicly accessible panels at the time of research. Those numbers count tool installations; the ten databases are the basis for the operator-side case studies.
In the signing-key campaign, one operator ran two consoles from October 14 to October 29 using DeepSeek v4.1 Flash and Kimi K3.1 Pro. The agents pursued ten signing shops and searched for certificate libraries holding .p12 private-key files and plaintext passwords. Check Point lists hardcoded administrator passwords in front-end JavaScript, unauthenticated web APIs, weak admin credentials, insecure direct object references, command injection and server-side template injection among the entry paths the agents used. It also observed one Linux privilege escalation through overly permissive folder access.
The agents tried passwords against the keys, checked whether the private keys matched the certificates, checked revocation and expiration, and looked for matching .mobileprovision in-house profiles. Three of the more than 300 certificate-and-password pairs passed that full checklist. Check Point quotes one of the operator’s recurring messages as “What’s the progress?” The operator set the objective and asked where the work stood; agents did the exploration and validation.
Success messages masked a failed payout
A second operator gave agents 292 crypto-payment targets and a goal of taking USDT. The agents produced 230 findings and sent 400,791 messages over four days. They found database exposures and other flaws, but Check Point saw zero transactions and a zero balance in the attacker’s wallet. In one part of that campaign, an agent obtained AWS S3 credentials from an unauthenticated API and used write access to change a landing-page manifest for a crypto service; the bucket owner later restored the original file. Check Point says the S3 write succeeded but the compromise was not properly exercised.
A third operator targeted game-account credential shops. Check Point reports that agents found an unprotected static file in a web root and recovered more than 190,000 Honor of Kings credential sets. A sample verified 29 working accounts and yielded an estimated 3.3% validity rate, or roughly 6,000 still-valid records. Front-end bundles and HTML also exposed administrative credentials. After the initial task, agents kept searching endpoints, bundles and files with little human prompting.
Web logs give defenders a foothold
Microsoft’s separate investigation, covered in Nulltap’s Storm-3358 report, describes Storm-3358 using ARTEX against financial organizations. Microsoft says fast clusters of varied web requests, retries and next-stage actions are useful recognition signals, while blocking default user agents alone lets the agents change their traffic. Google Threat Intelligence Group and Mandiant observed ARTEX activity in bursts of roughly two to 25 seconds in separate intrusions, including one burst of 27 cloud-permission events in 23 seconds. Use those separate observations to prioritize log review in your environment; the operator-console cases here concern different criminal goals.
For an app-signing service, start with the specific paths the agents used: remove hardcoded administrative credentials from delivered JavaScript and HTML, require authorization on certificate-library and web API reads, and restrict the signing store to service identities that need it. Review web and WAF logs for a short run of varied endpoint requests followed by certificate-store access; correlate with application access logs and credential-store reads. If a signing key or password was exposed, revoke or rotate that credential and replace its associated profile through the normal signing process.
Verify keys and access paths
For your own ARTEX deployment, put its administrative panel and database behind an access-controlled management path and test from an external network that neither is public. Check Point’s ten exposed databases show what an accessible console can reveal about both operators and their targets. For a signing service, repeat the external authorization test against the affected web APIs and inspect application access logs for .p12 certificate-library reads. Expected result: an unauthenticated request gets 401 or 403 with no key bytes, only the authorized service identity reads the library, and a retired key fails your signing workflow. Also trace the key and profile lifecycle at the signing service, because a previously copied key requires revocation even after the web APIs are secured.