Unverified50% confidenceBenchmarkExact time
An Anthropic red-team evaluation called out Zhipu GLM-5.3, stating it can autonomously write complete exploit code, with a success rate of up to 100% in bypassing its own defenses, and that its safety is lower than Claude models at the same tier.
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/30/2026
First Seen
Valid until: 12/29/2026
Sources
Related Entities
Related Claims
Unverified智能体安全应通过攻击成功率、过度拒绝率、攻击类型分类准确性三大核心指标量化评估74% similarUnverified开放权重使红队研究者能够进行白盒攻击测试，效率远高于只能通过API进行的黑盒测试68% similarVerified红队测试是AI安全领域的核心实践，红队成员系统性地尝试越狱自家模型以发现安全护栏漏洞68% similarUnverifiedIn adversarial testing, synthetic attacks still have approximately a 7% miss rate against Claude Code's safety classifier.68% similarUnverifiedSecurity researchers say Claude Code's default auto mode can be hijacked by prompt injection in web summary requests, thereby triggering code execution, with small-sample tests showing an attack success rate as high as 60% to 80%.68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/961338API
curl https://kongchang.com/api/v1/knowledge/claims/961338MCP
get_claim(id=961338)