Skip to content

Translation in progress Some entries still include text in the original language.

Dangerous capabilities · News

Oct 13 · Tuesday

Dangerous capabilities · News

Latest publications and developments
Oct 2Fri
  1. Google DeepMind, Help Net Security AI · 3 sourcesGoogle DeepMind · 3 sources · 4 reports · 43

    Google DeepMind releases SynthID Bio protein watermarking technology

    Google DeepMind releases SynthID Bio, bringing watermarking technology to synthetic biology by embedding imperceptible signatures directly into biological code, so the watermark can be verified both on digital models and on synthesized physical proteins while preserving their biological function. The method adapts its strategy by data type: fine-tuning amino acid choices for sequences and adjusting atomic coordinates for predicted 3D structures. The team paired the AlphaProteo design method with a ProteinMPNN sequence generator with SynthID Bio, ran experiments on targets such as VEGF-A, the SARS-CoV-2 spike protein receptor-binding domain, and PD-L1, and verified through wet-lab experiments that the watermark does not impair protein function. DeepMind says this is a world first, embedding signatures into protein sequences without affecting biological function. Additionally, DeepMind discussed in a podcast the role of provenance and how imperceptible watermarks can protect digital media and biological designs.

    Latest developments
    1. Google DeepMind podcast introduces SynthID watermarking technology

    2. Google DeepMind releases SynthID Bio protein watermarking technology

  2. Goodfire ResearchGoodfire Research · 46

    Goodfire 发布基于蛋白质嵌入的生物安全监控器

    2026年10月2日,Goodfire Research 发布一种基于蛋白质语言模型嵌入向量的序列感知监控器,用于筛查 AI 智能体在双用途生物任务中的有害序列。该方法同时利用序列本身和任务上下文,在自建基准上优于前沿模型护栏,且在双用途任务上拒答更少。监控器对改写和片段化等对抗攻击更稳健:当原始蛋白存在于筛查数据库时 AUROC 约为 0.98;原始蛋白不在库中时各方法性能均下降,但该方法在低误报率下仍保持优势。其泛化能力还扩展到留出的蛋白家族、酶类毒素机制留出集以及外部细菌毒素数据。

Oct 1Thu
  1. Goodfire · 28

    Goodfire 推出基于蛋白质模型嵌入向量的生物安全序列感知监控器

    Goodfire研究人员提出基于蛋白语言模型表征的序列风险监控器,用于生物研究智能体工作流中的双重用途筛查。他们在自建基准上发现,所测前沿模型的拒答主要受任务描述影响,对危险和良性序列的区分不足;序列监控器在被测改写、片段化和部分留出条件下优于序列比对筛查。作者报告该方法达到毫秒级处理,但重设计蛋白是否保留生物活性仅通过计算估计,未经实验验证。结果限定于被测序列与基准,不能视为全面生物安全保证。

Jul 16Thu
  1. Google DeepMind · 29

    Google DeepMind 与 Isomorphic Labs 联合公布生物韧性方案

    Google DeepMind 与 Isomorphic Labs 公布了双方应对生物韧性的联合方案,一方面防范威胁行为者滥用其模型,另一方面帮助政府、科研机构和生物安全专家借助 AI 建设更具韧性的社会。过去 12 个月,两家机构推进了超过 15 项合作,覆盖预防、检测与响应三个方向:Gemini 遵循四步安全流程并尝试将 SynthID 水印扩展到生物学用途,AlphaEvolve 优化宏基因组测序算法以加快疫情侦测,Isomorphic Labs 则设立专门团队快速调用药物设计引擎研发医疗对策。该工作属于其对 CBRN 风险的整体管理的一部分,并与 Frontier Safety Framework 相衔接。

Dec 9, 2025Tue
  1. Jan Leike · 34

    Anthropic 安全缓解措施漏洞赏金计划

    Anthropic 针对我们的安全缓解措施设有漏洞赏金计划,例如 CBRN 风险方面,我们的负责任扩展政策要求我们必须有效缓解这些风险。 如果你对此感兴趣,请报名参加!你可以通过攻破我们的防御来帮助 AI 安全并赚取报酬。👇

You have reached the end