deepseek-v4.1IMAGE
Web pages can steer a DeepSeek V4.1 agent by quoting its control tokens
The author found that a web page can steer a DeepSeek-V4.1 agent without a jailbreak simply by quoting the model's own control tokens, and fixed it in the native engine running BLACKFOLD.
BlackwellBoy
@BlackwellboyArchived 10 days agoOriginally posted 2026.10.22
BlackwellBoy @BlackwellboyOriginally posted 2026.10.22
Found a way for a web page to steer a DeepSeek-V4.1 agent, and fixed it in the native engine I run BLACKFOLD on. The page doesn't need a jailbreak. It just needs to quote the model's own control tokens. 🧵
x.com/blackwellboy/status/2113044818746314938Archived 2026.10.22 11:58