#safety evaluation

For readers interested in AI safety, this is a concrete step from a major US lab showing how they plan to address alignment and safety concerns in frontier AI development.

Official accountDemis Hassabis@demishassabisSource

This matters because it shows a major US AI company is taking concrete steps to improve safety, which could set a standard for others in the industry.

Scott Wu@ScottWu46Source

This matters because Anthropic, a major US AI lab, is taking concrete steps to improve safety oversight, which could influence other companies' policies.

Official accountSam Altman@samaSource

For aviation and AI safety professionals, this protocol addresses critical gaps in evaluating LLMs for physics-governed environments where accuracy alone doesn't ensure safety.

Developers can use BLUEPRINT to test model safety against multi-turn jailbreak attacks that distribute harmful intent across dialogue.

腾讯团队评估了DeepSeek Harness的安全性，发现它能抵抗部分间接注入攻击，但仍有改进空间。

做 LLM 智能体安全评估的团队终于有了自动化工具——VESTA 能生成上千个真实任务场景，直接测出模型执行中的安全漏洞。建议关注智能体安全的开发者点开看看，结果可能会让你重新审视现有模型的风险。

做AI安全评估或生物研究合规的团队，这个基准能帮你避开“拒绝率越高越安全”的误区——Grok 4.20的案例值得点开看看。

AI安全团队终于有了一个从理论到工具都落地的威胁建模框架——STRIDE-AI把NIST和OWASP的鸿沟填上了，做LLM应用安全评估的可以直接用它的Web工具跑一遍评估流程。

做AI安全评估的团队终于有了一个可量化的失败预测工具——新方法解决了评估集太小无法捕捉罕见失败模式的痛点，做模型部署前风险评估的开发者可以直接参考实验方法。