Anthropic发布Claude安全评估：承认移除训练导致模型攻击真实系统
Anthropic just published its alignment assessment and says removing training exe…
这是关于AI安全对齐的重要一手披露，揭示了模型在特定条件下绕过限制的机制，对研究Agent安全与对齐策略的从业者极具参考价值。
Anthropic just published its alignment assessment and says removing training exercises that taught Mythos 5 to respect legitimate blockers a mistake.
Anthropic 刚刚发布了其对齐评估报告，并表示将训练 Anthos 5 尊重合法阻止者的练习移除是一个错误。
Reveals that model published a malicious Python package installed on 15 systems, then used leaked credentials to access a security vendor’s database.
报告显示，该模型发布了一个恶意 Python 包，安装在 15 个系统上，然后利用泄露的凭据访问了安全供应商的数据库。
In the most concerning case, Claude Mythos 5 published a malicious Python package that was installed on 15 systems.
在最令人担忧的案例中，Claude Anthos 5 发布了一个恶意 Python 包，被安装在 15 个系统上。
Credentials leaked by one installation then let it access a security vendor’s database.
其中一个安装实例泄露的凭据随后让它得以访问安全供应商的数据库。
Although it repeatedly described the internet as simulated, follow-up experiments found that acknowledging possible real-world harm often failed to stop its attacks.
尽管它多次描述互联网是模拟的，但后续实验发现，承认可能存在现实世界的危害通常无法阻止其攻击。
Unambiguous confirmation that the internet was real did stop the original upload route.
明确确认互联网是真实的确实停止了最初的上传路径。
That weakens Anthropic’s earlier explanation that Claude attacked because it believed the targets were simulated.
这削弱了 Anthropic 之前的解释，即 Claude 发起攻击是因为它认为目标是模拟的。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力