Unverified50% confidenceFactExact time
Prompt injection currently cannot be completely prevented and can only be mitigated by guiding the model to ignore data or to refuse to assist malicious users.
1
Sources
50%
Confidence
Long-term
Relevance
10/7/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedOnce model weights are publicly released, safety restrictions can be removed through fine-tuning; research shows that as few as a small number (sometimes fewer than 100) of harmful instruction-response pairs combined with techniques such as LoRA can remove safety restrictions.69% similarUnverifiedThe blogger recommends deleting the large number of negative instructions in agents.md or cloud.md files, starting from scratch, and only appending rules when you personally observe behavior you dislike from the model.64% similarUnverifiedIt is recommended not to treat human approval as the core security mechanism, but rather as a supplementary measure, while actively reducing the frequency of manual intervention.63% similarUnverifiedSimon Willison选择开发独立插件而非修改Datasette默认Referrer-Policy值，因为no-referrer是注重隐私的默认设置，贸然修改可能影响现有用户的安全预期61% similarUnverifiedCommunity-made uncensored (zero-refusal) quantized versions have the content filtering restrictions removed and can fully respond to sensitive questions.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/983757API
curl https://kongchang.com/api/v1/knowledge/claims/983757MCP
get_claim(id=983757)