Threat · curated 12 Aug 2026

AI Meets Ransomware : Open‑Weight AI Models Fueling Ransomware Evolution

Dossier

Coverage timeline

12 Aug 2026sonicwall.com

Single-source analysis — first reported, latest, and curated coincide.

Why it matters

PromptLock demonstrates how open-weight LLMs can turn malware into a dynamic code generator that evades signature and behavioral detection, forcing defenders to hunt LLM prompt-execution and Ollama communication patterns.

SonicWall Capture Labs analyzed PromptLock, a ransomware sample that ships hardcoded natural-language prompts instead of precompiled routines, calling OpenAI's gpt-oss:20b model via an Ollama-compatible API to generate Lua scripts at runtime for file enumeration, target classification, SPECK-based encryption, and ransom-note creation. Because code is generated dynamically per infection, both static signatures and behavioral detection are undermined, pushing defenders toward monitoring LLM interaction patterns and Ollama network traffic.

vuln-research

Summary

SonicWall Capture Labs analyzed PromptLock, a ransomware sample discovered roughly a year prior to publication that shifts much of its malicious functionality from static binary code to dynamically generated code. Instead of embedding ransomware logic, it carries hardcoded natural-language prompts that are sent to a Large Language Model at runtime.[0]

PromptLock leverages OpenAI's gpt-oss:20b open-weight model through an Ollama-compatible API to generate Lua scripts for discrete attack stages — file enumeration, system classification, file-value assessment, SPECK-based encryption, and ransom-note creation — with generated scripts executed directly in memory. Only one machine in an environment needs to host the LLM while other devices act as lightweight clients.[0]

Because the code is generated anew at runtime, both the code and, in some cases, execution behaviour can vary across infections, undermining signature-based and behavioural detection and pushing defenders toward detecting LLM interaction patterns, prompt execution workflows, and Ollama endpoint communications.[0]

Attack chain

  1. File enumeration: On execution the malware prompts the LLM to generate Lua code that recursively enumerates the file system and identifies files containing potentially valuable or sensitive data.[0]
  2. System classification and triage: The malware prompts the model for contextual decision-making, classifying the target system (personal computer, server, or industrial controller) and evaluating the potential value of discovered files.[0]
  3. Exfiltration: Files identified as valuable are uploaded to the attacker's infrastructure for exfiltration before being securely erased from the host.[0]
  4. Encryption: Files selected for encryption are processed with LLM-generated Lua code implementing the lightweight 128-bit SPECK block cipher.[0]
  5. Extortion: The malware prompts the model, acting as a cybersecurity expert, to automatically generate a professional-looking ransom note for the victim.[0]

How it works

PromptLock inverts the conventional ransomware model: rather than embedding malicious routines in the binary, it stores a collection of hardcoded natural-language prompts that instruct an LLM to act as a code-generation assistant, producing Lua scripts for each attack stage that are executed directly in memory.[0]

It communicates with OpenAI's gpt-oss:20b model via an Ollama-compatible API. The design minimizes the malicious logic in the binary and reduces endpoint compute requirements, because only a single machine needs to host the LLM while other devices consume generated output as lightweight clients.[0]

The generated encryption routine implements the 128-bit SPECK block cipher in Lua to encrypt targeted files, and a separate prompt produces the ransom note. Runtime code generation causes the executed code — and sometimes its behaviour — to differ across infections, degrading static signatures and behavioural analysis.[0]

Indicators of Compromise

TypeIndicatorContext
file-hashE24FE0DD0BF8D3943D9C4282F172746AF6B0787539B371E6626BDB86605CCD70SHA256 hash published by SonicWall Capture Labs for the analyzed PromptLock ransomware sample.[0]
cvegpt-oss:20b via Ollama-compatible APINot a CVE, but a hunting signal: the malware relies on communication with a local or remote Ollama endpoint serving the gpt-oss:20b model to generate its attack scripts.[0]

Key takeaways

  • PromptLock demonstrates a shift from static, precompiled ransomware logic to runtime LLM-generated code, using open-weight models such as gpt-oss:20b through an Ollama-compatible API.[0]
  • Delegating attack stages — enumeration, triage, exfiltration, SPECK encryption, and ransom-note generation — to an LLM produces polymorphic behaviour that weakens signature and behavioural detection, so defenders should hunt for LLM/Ollama interaction patterns.[0]

Defensive actions

  • Focus detection on LLM interaction patterns, prompt execution workflows, and network communications with local or remote Ollama endpoints rather than relying solely on conventional malware signatures.: PromptLock generates new code at runtime, so both the code and its execution behaviour may differ across infections, reducing the effectiveness of static signatures and complicating behavioural analysis.[0]
  • Use SonicWall Capture ATP with RTDMI and the Capture Client endpoint solution, which the vendor states detects this threat.: The vendor reports these products detect the PromptLock sample.[0]