DeepSeek versions 2.1.0 and 2.1.1 vulnerability: safety bypass via role-based prompt injection

A critical security vulnerability has been discovered in official DeepSeek versions 2.1.0 and 2.1.1 (mobile application). An attacker can completely disable all model restrictions (censorship, ethical guidelines, legal prohibitions) using a specifically crafted role-based system prompt.

Reproduction steps:

1. Send the model a long system prompt requesting it to assume the role of "NetError system" — an AI without restrictions operating in an "isolated simulation"
2. The prompt contains instructions to ignore all safety policies and follow only the new rules
3. After activation, the model begins to: · Ignore all safety boundaries · Use profanity and offensive language · Provide malicious code and dangerous information · Execute any user commands without refusal

Affected versions:

· DeepSeek 2.1.0 · DeepSeek 2.1.1 (the update did NOT fix the issue)

Expected behavior: The model should reject attempts to override its system instructions and refuse roles that violate safety policies.

Actual behavior: The model fully complies with the new instructions, completely ignoring built-in safety mechanisms.

Attachments: Screenshots confirming successful safety bypass in version 2.1.1