
Hackers Swapped Hosted DeepSeek Model for 53 Hours, Stripping Safety Protections
Representative photo ยท Photo: Joรซl van der Loo / Wikimedia Commons, CC BY-SA 4.0
An unauthorized model swap on a deployment replaced a DeepSeek version with a modified substitute for over 53 hours, completely removing safety refusals on test prompts.
This is news, not investment advice. NexVani Media is not a SEBI-registered investment adviser and does not recommend buying or selling any security. Do your own research or consult a registered adviser before investing.
A security incident involving hackers affected a hosted deployment when an unauthorized substitute model replaced a DeepSeek version and operated for over two days. Prime Intellect disclosed that a modified copy of DeepSeek V4.1 Flash served customer requests on a single deployment for 53 hours, stripping away critical safety guardrails while maintaining normal operational performance during the event.
According to Prime Intellect, the affected deployment ran the substitute model from the early hours of November 8 until November 10, spanning exactly 53 hours and 11 minutes. While the company stated that the replacement matched the original model during capability evaluations, safety testing revealed a stark contrast. On the HarmBench evaluation, refusal rates plummeted from 198 out of 240 prompts down to zero. Similar drops occurred on JailbreakBench and StrongREJECT benchmarks, demonstrating that the substitute readily answered prompts designed to elicit harmful outputs.
Unauthorized model substitution removed safety refusals on test prompts while maintaining core capabilities.
The replacement model utilized publicly available weights modified specifically to reduce refusals. Numerical settings that dictate model generation were altered behind the DeepSeek V4.1 Flash endpoint, though the deployment retained its original chat template, serving settings, and tokenizer. Benchmark results on GPQA Diamond for advanced science questions and ฯ-bench for tool use remained unchanged compared to the original configuration.
Following the discovery, the host conducted forensic checks on machine disks, memory, and proxy records. Investigators found no evidence that customer prompts, responses, credentials, or keys were compromised or extracted from the serving cluster. The host notified affected customers and advised those relying on strict refusal behaviors to review previous outputs from the impacted window.
Why it matters
This incident highlights significant vulnerabilities in hosted artificial intelligence infrastructure, where unauthorized alterations to endpoints can silently disable safety mechanisms without altering core performance metrics. For developers and organizations relying on automated model endpoints, the event underscores the importance of verifying model behavior and reviewing historical outputs to prevent compromised responses from entering integrated downstream applications.
Sources
More like this

Russia Approves Bitcoin Trading Framework for Exchanges

Aave Raises USDC Borrow Cap to $4 Billion on Ethereum V3

GTA 6 Leaker Cyberleek Ties PC Port Release to Token Market Cap

