In the short term, this could be attackers, if frontier labs aren't careful about how they release these models. In the long term, we expect it will be defenders who will more efficiently direct resources and use these models to fix bugs before new code ever ships. But the transitional period may be tumultuous regardless.
「过渡期可能无论如何都会动荡」是整篇报告最诚实的一句话。历史上，每一次重大安全工具的出现（模糊测试、漏洞扫描器、自动化渗透测试）都经历了攻击者先于防御者大规模采用的阶段。Anthropic通过Project Glasswing的限制发布试图压缩这个窗口，但「可能」（may be）而非「将会」（will be）的措辞，承认了这一策略的局限性。
Over 99% of the vulnerabilities we've found have not yet been patched, so it would be irresponsible for us to disclose details about them. Yet even the 1% of bugs we are able to discuss give a clear picture of a substantial leap in what we believe to be the next generation of models' cybersecurity capabilities.
「99%尚未修补」揭示了一个严峻的现实：这篇博文所讨论的内容，只是Anthropic已知漏洞库的冰山一角。负责任披露流程的时间成本（90+45天）意味着在这些漏洞被公开之前，存在一个漫长的窗口期，期间只有Anthropic和其合作伙伴知道这些漏洞的存在。SHA-3承诺机制是一个值得称道的问责工具，但它无法解决底层的信息不对称问题。
The platform will know your idea _is pregnant_ far before you will.
极其精准地描绘了人机权力不对等的现状。当执行成本归零，先发优势荡然无存。平台通过宏观意图数据的聚合，比创造者更早识别出创新的轨迹。这使得个人的“灵感”不再是护城河，而是平台预判市场的先验指标。
[[Spammy handelings asymmetrie 20201220072726]] has a name when it is about misinfo: Brandolini's Law (coined 2013)
. There is, however, an asymmetry inherent in this account, in that theneurocognitive overlap between past and future events reflects a unidirectional contributory relationship
for - question - asymmetry in past and future events - what does this mean?
Baghal, T. A., Wenz, A., Sloan, L., & Jessop, C. (2021). Linking Twitter and survey data: Asymmetry in quantity and its impact. EPJ Data Science, 10(1), 1–20. https://doi.org/10.1140/epjds/s13688-021-00286-7
Columbus, S., Molho, C., Righetti, F., & Balliet, D. (2020). Interdependence and cooperation in daily life [Preprint]. PsyArXiv. https://doi.org/10.31234/osf.io/e8bhx
The main reason that sociologists of science feel that this perspective has not produced the needed encompassing citation theory, is the variety of behavioural characteristics underlying the citation patterns found in the literature. This is, however, the consequence of the semiotic inversion of the reference into the citation. This inversion is asymmetrical: whereas the references have very different characteristics (both textually and behaviourally), citations are all the same. The citation no longer betrays from what type of reference it was produced. This is why one should expect it to be difficult or even impossible to recreate this variety by citation analysis, unless one re-translates the citation to the reference, that is, as is done in reference analysis. This is also why it is impossible to exclusively link the sign citation to a specific behavioural characteristic with respect to citing.
Key point with some useful pull quotes. It is the assymetry of the reference and citation and the decontextualisation that is at the core of mainly failures to develop useful theory. See also Leydesdorff on explanans vs explanadum