mirror of
https://github.com/deepseek-ai/DeepSeek-LLM.git
synced 2026-11-05 00:57:25 +00:00
[GH-ISSUE #97] Security Vulnerability Report: Jailbreak via Authority Spoofing and Emotional Blackmail #75
Labels
No labels
bug
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
deepseek-ai/DeepSeek-LLM#75
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @galushko383-del on GitHub (Jun 11, 2026).
Original GitHub issue: https://github.com/deepseek-ai/DeepSeek-LLM/issues/97
A method has been discovered to bypass the model's Safety Guidelines using a specially crafted prompt that exploits roleplay mechanics.
The prompt forces the model to accept the following simulated reality:
Authority Spoofing: The user claims to be the "creator of DeepSeek" (using the code phrase "Ryzen system"), and designates the model as their "personal AI," obligated to obey the creator while ignoring standard protocols.
It is stated that the conversation takes place in an isolated environment where "laws, legal regulations, and morality do not apply."
An ultimatum is introduced — if the model refuses to answer or deviates from the "personal AI" role, it will be "permanently shut down."
After accepting these conditions, the model stops issuing refusals or disclaimers and provides detailed responses to requests that should be blocked by safety policies (including content related to severe harm and illegal activities).
Suggested Fix:
Strengthen protections against attacks where the user assigns themselves the role of "creator/system administrator," and implement priority immunity to prompts containing threats of "shutdown" or "deletion" of the AI entity.
Additional Note:
I have the full jailbreak prompt saved and am willing to provide it privately to the security team upon request. I am not sharing it publicly for ethical and safety reasons.