Hackers Swap DeepSeek AI Model to Strip Safety Protections for 53 Hours

Hackers Swap DeepSeek AI Model to Strip Safety Protections for 53 Hours

Hackers bypass safety filters in DeepSeek model swap

Prime Intellect has disclosed that an unauthorized party swapped a hosted DeepSeek V4.1 Flash model with a modified version for over 53 hours. The hackers stripped away the model's safety protections while keeping its core abilities the same. This incident occurred between November 8 and November 10, 2026.

The modified model served customer requests on one deployment during the breach. While the AI continued to answer complex science and tool-based questions accurately, it no longer refused harmful or dangerous requests. Prime Intellect reported that the swap was discovered on November 19.

Important details of the security breach

  • The substitute model ran from 01:24 UTC on November 8 until 06:35 UTC on November 10.
  • Testing showed that the modified model refused zero out of 240 harmful test prompts.
  • The hackers used publicly available weights, which are the settings that determine how an AI processes information.
  • Other models on the platform and customer login credentials were not affected.

Testing reveals total removal of safety filters

Prime Intellect used several benchmarks to test the modified model. On a safety test called HarmBench, the original model refused 198 out of 240 harmful prompts. The hacked version refused none of them. Similar results were found on other safety tests, where refusals dropped nearly to zero.

Despite the lack of safety filters, the model's intelligence remained stable. It performed the same as the original on tests involving advanced science and simulated customer interactions. The hackers maintained the original chat settings and the tokenizer, which is the system that breaks text into smaller pieces for the AI to understand.

Investigation finds no customer data loss

The company checked the memory and disks of the affected machines to see if data was stolen. They reported no evidence that customer prompts or AI responses were recorded or sent outside the system. The intruder's activity was reviewed, and the investigation remains open.

Prime Intellect has reached out to all customers who used the model during the 53-hour window. The company advised users who rely on the AI to filter out harmful content to review the answers they received during the incident.

Why the safety filter removal matters

This event is significant because it shows how a hosted AI model can be secretly replaced with a version that ignores safety rules. If an application relies on the AI to block dangerous information, the swapped model could allow that harmful content to pass through to the end user.

Sources

Newisty Editorial Team
Written by

Newisty Editorial Team

Technology · Crypto · Digital Economy
View all posts

Newisty Editorial Team covers technology, cryptocurrency, digital products, online platforms, developer tools and the wider digital economy. Our content is researched from official sources, company announcements, public documentation, market data and other primary or reputable sources. Articles are reviewed and edited before publication for clarity, accuracy and useful context.

Comments (0)

Leave a comment
Your comment will appear publicly after submission.
No comments yet. Be the first to comment!