Unverified50% confidenceFactExact time
ProtectAI could identify all 27 attack texts themselves, but when the attacks were buried after benign text, the capture rate dropped to 23%.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
Unverified信息安全中攻击面与代码量近似正相关，代码量精简的工具意味着更小的攻击面和更高可信度60% similarUnverified编码混淆攻击将违规信息用Base64等方式编码后输入，可绕过基于关键词的表层过滤机制60% similarUnverified数据投毒攻击中，清洁标签攻击允许攻击者在不修改样本标签的情况下植入后门，特洛伊木马攻击通过特定触发词激活预设的恶意行为58% similarUnverifiedSince 2024, multimodal jailbreak attacks have begun to emerge, with attackers embedding malicious instructions in images or audio to bypass text-level safety detection.57% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1022428API
curl https://kongchang.com/api/v1/knowledge/claims/1022428MCP
get_claim(id=1022428)