Meta Muse 的安全设计引发讨论
作者认为 Muse 是首个在安全上投入较多的个人 agent，但指出分类器强度与体验之间存在取舍，并提到 Meta 漏洞赏金计划。
this is probably the first personal agent released with this much attention to security, unlike most of the others. the problem with personal agents is that even if you sandbox them properly, you’re still giving them access to all your third-party services. unlike humans, agents are notoriously susceptible and naive to social-engineering attacks through prompt injection, which can trick them into exfiltrating information or performing actions through connected services. meta blog says muse uses multiple classifiers to mitigate these risks, but not sure how good they are the problem is if you make classifier too strong it brings bad ux, so there is a tradeoff between security and ux. meta’s security team is quite strong and it has a solid bug bounty program, so this looks like a good opportunity to farm some bounties.
这条是引用着另一条写的：「Meta 为 Muse 漏洞悬赏最高 30 万美元」