Unverified50% confidenceFactExact time

OpenAI 和 Anthropic 的研究承认模型安全护栏在面对精心设计的越狱提示时仍存在被绕过的可能性

1

Sources

50%

Confidence

Medium-term (~90 days)

Relevance

8/8/2026

First Seen

Valid until: 11/6/2026

Sources

AI Company Uses Social Engineering to Attack Open Source Maintainers and Inject Malicious Code: A Supply Chain Security Wake-Up Call

Related Claims

UnverifiedOnce full weights are made public, anyone can remove safety guardrails through fine-tuning, which is one of the core reasons some top-tier models do not release their weights.70% similar UnverifiedExperience following Meta's release of LLaMA weights indicates that open weights may be used for fine-tuning to remove safety guardrails69% similar

Cite This Claim

Stable URI

https://kongchang.com/claim/707568

API

curl https://kongchang.com/api/v1/knowledge/claims/707568

MCP

get_claim(id=707568)