If public models can already do useful work inside that kind of workflow, then the story is not 'Anthropic has a magical cyber artifact.' The story is that serious AI-assisted vulnerability research is no longer confined to a single frontier lab.
这一发现挑战了Anthropic试图构建的叙事：即高级AI安全研究需要受限访问。研究表明，公共模型已经能够复制关键的安全发现，这意味着真正的'护城河'不是模型访问，而是验证、优先排序和操作化的能力。这打破了'只有前沿实验室才能进行高级AI安全研究'的神话。
Introduction: AI is now recently everywhere but we still need humans
Betsch, C., & Böhm, R. (2016). Detrimental effects of introducing partial compulsory vaccination: Experimental evidence. European Journal of Public Health, 26(3), 378–381. https://doi.org/10.1093/eurpub/ckv154
Amidst the global pandemic, this might sound not dissimilar to public health. When I decide whether to wear a mask in public, that’s partially about how much the mask will protect me from airborne droplets. But it’s also—perhaps more significantly—about protecting everyone else from me. People who refuse to wear a mask because they’re willing to risk getting Covid are often only thinking about their bodies as a thing to defend, whose sanctity depends on the strength of their individual immune system. They’re not thinking about their bodies as a thing that can also attack, that can be the conduit that kills someone else. People who are careless about their own data because they think they’ve done nothing wrong are only thinking of the harms that they might experience, not the harms that they can cause.
What lessons might we draw from public health and epidemiology to improve our privacy lives in an online world? How might we wear social media "masks" to protect our friends and loved ones from our own posts?
Vu, Jonathan T, Benjamin K Kaplan, Shomesh Chaudhuri, Monique K Mansoura, and Andrew W Lo. ‘Financing Vaccines for Global Health Security’. Working Paper. Working Paper Series. National Bureau of Economic Research, May 2020. https://doi.org/10.3386/w27212.
social engagement, public knowledge, and the mission of promoting enlightenment and critical inquiry in society