Exposed Server Unveils AI-Powered Phishing Kit Fueling WebDAV Malware Attacks

Introduction

A spectacular operational security failure by a threat actor has handed defenders an unprecedented look inside the modern phishing supply chain. Researchers at Rapid7 discovered an unprotected server hosting 1,048 files that document the end-to-end creation of AI-enhanced phishing campaigns, complete with live operations targeting Windows users in Mexico through fraudulent government identification portals. The cache includes everything from prompt engineering experiments for lure generation to WebDAV-based payload delivery mechanisms, offering a rare blueprint of how cybercriminals are operationalizing generative AI for social engineering at scale.

Anatomy of an Exposed Attack Infrastructure

The compromised server functioned as a development staging ground rather than a command-and-control node, which explains the breadth of artifacts preserved. Rapid7's analysis reveals a methodical workflow: the operator maintained separate directories for lure templates, filename obfuscation testing, execution validation logs, dropper variants, and handwritten builder notes. Two distinct campaign chains were documented, one in active deployment and another in late-stage preparation. This level of documentation suggests either a solo operator with rigorous habits or a small team treating phishing as a software engineering discipline.

Generative AI as a Force Multiplier for Social Engineering

Perhaps the most significant revelation is the extensive use of large language models to craft convincing government-themed lures. The toolkit contains dozens of prompt iterations showing the operator refining Spanish-language content for Mexican Instituto Nacional Electoral (INE) and CURP (Clave Única de Registro de Población) impersonation pages. These aren't crude translations—the prompts specify tone, formatting, legal terminology, and even regional dialect nuances. The operator appears to have used AI to generate multiple lure variants per campaign, then A/B tested them against target demographics. This represents a fundamental shift from the spray-and-pray phishing of previous years toward precision social engineering informed by data science.

WebDAV: The Overlooked Delivery Vector

The active campaign leverages WebDAV (Web Distributed Authoring and Versioning) for payload delivery, a technique that bypasses many traditional email security controls. Victims receive a link appearing to lead to a government document portal; clicking it mounts a remote WebDAV share that presents a malicious LNK or VBScript file masquerading as a PDF. Because WebDAV traffic often traverses port 443 and resembles legitimate cloud storage synchronization, it evades network inspection rules tuned for executable downloads. The toolkit includes extensive testing logs showing the operator validating this technique across Windows 10 and 11 versions, with specific attention to Mark-of-the-Web bypasses and SmartScreen reputation evasion.

Infostealer Payload Targeting Mexican Financial Data

The final stage deploys a customized infostealer configured to harvest browser credentials, cryptocurrency wallet data, and—critically—Mexican banking application tokens. Builder notes indicate the operator prioritized stealth over persistence, using reflective DLL injection and living-off-the-land binaries (LOLBins) like mshta.exe and regsvr32.exe. The malware exfiltrates via encrypted HTTPS POST requests to a separate infrastructure tier, suggesting compartmentalization between delivery and collection layers. Rapid7 has shared indicators of compromise with Mexican CERT and financial sector ISACs.

Operational Security Lessons from the Operator's Mistakes

• No authentication on staging infrastructure: The server required no credentials, allowing full directory listing and file download.
• Retention of development artifacts: Failed experiments, deprecated templates, and internal notes were never purged.
• Reused infrastructure patterns: File naming conventions and directory structures match previous campaigns attributed to the same threat actor.
• AI prompt history preserved: The operator saved entire conversation logs with the LLM, revealing intent and iteration logic.

Conclusion

The Rapid7 find serves as a stark reminder that threat actors are not immune to the same operational hygiene failures that plague legitimate organizations—and when they slip up, the intelligence windfall can be transformative. The 1,048-file archive provides a Rosetta Stone for understanding how AI-assisted phishing kits are assembled, tested, and deployed against specific populations. For security practitioners, the actionable takeaways are clear: audit WebDAV egress policies, deploy endpoint detection capable of identifying LNK execution from remote shares, and invest in phishing-resistant authentication (FIDO2/WebAuthn) that renders stolen credentials useless. As generative AI continues to refine the attacker's quill, the only sustainable defense is a detection architecture that assumes the lure will be perfect—and focuses instead on neutralizing the hook. Grivyonx Cloud's unified AI automation and cybersecurity platform helps organizations implement precisely this kind of behavior-centric resilience across hybrid environments.