Large language models have drastically changed software development, and cybercriminals are quickly adopting these tools to build malicious payloads.
While vendors like OpenAI and Anthropic have implemented strict cyber-safety guardrails to block harmful requests, other models present a different landscape.
A recent analysis by Check Point Research highlights how DeepSeek models, which often have lower refusal rates for harmful prompts, are enabling threat actors to generate complex, end-to-end malware with minimal expertise.
Researchers analyzed nearly 3,000 files attributed to DeepSeek in public telemetry, identifying 1,383 malicious or dangerous samples. Among these, they discovered a highly unusual Python Flask application named InfernoGrabber.
This sample demonstrated a dangerous, previously unseen technique in the wild: in-browser ransomware.
Unlike traditional malware that requires a user to download a payload or exploit a system vulnerability, this threat operates entirely within the web browser by abusing legitimate features.
AI-Built Browser Ransomware Workflows
The attack centers on the File System Access API, a legitimate browser capability designed for web applications like photo editors and creative tools.
When granted permission, this API allows a website to read, write, and enumerate files in a local directory.
While browser engineers have previously documented the theoretical risk of ransomware abusing this API, the DeepSeek-generated sample turned this concept into a functional attack chain.
The victim is lured to a web page that appears to be a Discord avatar AI upscaler. To process the image, the site prompts the user to grant file-system access to a local folder.
Because users expect web applications to request permission to save edited files, this social engineering tactic feels natural and lowers suspicion.
Once the user clicks “Allow,” the malicious page can silently read, exfiltrate, and encrypt the files in the selected folder. Finally, it displays a ransom note demanding Bitcoin, all without ever installing a native executable on the device.
What makes this discovery alarming is how the AI model reasoned across existing knowledge. The attacker likely submitted a broad prompt asking for an all-in-one malware tool.
In response, DeepSeek hallucinated a Frankenstein-like scaffold of keyloggers and stealers. However, critically, it successfully connected the malicious intent to the real-world File System Access API.
This shows that frontier AI models are no longer just enhancing known attacker techniques; they are helping unskilled actors discover and operationalize novel attack vectors that bypass traditional endpoint protections, checkpoint said.
While this browser-native ransomware technique affects Chromium-based desktop browsers, it poses a severe risk to Android users.
Safari and Firefox do not expose the same file picker methods. However, recent Chrome for Android updates allow web applications to read and modify selected folders directly.
Follow us on Google News , LinkedIn and X to Get More Instant Updates. Set Cyberpress as a Preferred Source in Google.
