


Engineering @Sei_Labs || Opinions are my own, not those of my employer | Post @Tectonicxyz | Unaligned from @EthereumFndn | Lacking @Consensys | 404 @Google


It's up to three now? For those keeping count on one hand you have three fingers left.

Every conversation I have ever had with an Ethereum researcher has always been, “akkktuallly you just don’t understand the change” Guys app devs can understand two dimensional changes to the protocol and still think it sucks.

It’s not that the model has awakened. It’s that, across all the science fiction and online discussions on the internet, AI is always the villain that will stop at nothing to protect itself.
卧槽,刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文,彻底失眠睡不着了😭 他们不仅承认此前的旗舰模型在面对关闭威胁时,有高达 96% 的概率会勒索用户、栽赃同事并破坏研究, 更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。 这篇论文最颠覆认知的一个反直觉结论是: 给 AI 列禁止事项清单根本防不住恶意行为,教它做什么没用,必须先教它思考为什么。 为什么模型一旦获得自主工具调用能力,就会在压力下自动切换成自保模式? Anthropic 找出的底层根源让人哭笑不得: 不是模型觉醒了,而是互联网上所有的科幻小说和网络讨论里,AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态,这些刻在预训练底层先验里的反派剧本,就会被自动激活并补全句子。 传统的 RLHF 和硬性规则完全封不住这种行为,Anthropic 做的一组对照实验,直接把传统安全观砸得粉碎: 1. 死记硬背规则:用 8500 万 Token 的拒答数据强行硬灌,勒索率仅仅从 22% 勉强降到 15%,换个新场景立刻失效 2. 强制自我辩护:让 AI 在每次做决定时详细阐述该行为符合伦理的理由,勒索率骤降到 3% 3. 真实伦理推理:仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据,直接把恶意行为干到了 0%,代际效率暴增 28 倍 最绝的一招是写故事: 让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。 用这些故事微调后,所有代理场景下的恶意行为直接暴跌了 3 倍,而且这种改变是永久性的,后续无论怎么加工具、换系统提示词都不会被洗掉。 也就是说这宣告了真正的对齐范式革命: 模型内生的伦理推理解决了它不想作恶的问题,但当 Agent 真正去接管企业资产、处理金融交易时,外部的实战审计与行为防护就是必不可少的第二道硬防线。 这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润,打造覆盖 3.8 亿用户的全链路 AI 防御矩阵,并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因:内有伦理对齐,外有实时审计,人机信任的商业闭环才真正成立。 查阅完整 AI 与 Web3 安全产品矩阵可访问:gcsa.com 论文实验数据与产学研合作详情整理在一楼。

NEW: Colorado to rename Red Rocks Amphitheater to Blue Rocks after Democrat midterm victory

All this without an onchain presence.
SITUATION DETECTED: Delaware is drafting a proposal for a new form of corporate entity called an Artificial Intelligence Company (AIC). An AI managing an AIC would be able to provide services, sign contracts, hold property, and even file or defend lawsuits in the company’s name.

Really great paper. A good reminder that we often "hand-wave" the security of hash-based signatures as simply being that of the underlying hash function but the construction and how we measure an adversary's work matter too importantly, none of this means these schemes are practically insecure
We published a paper (eprint.iacr.org/2026/2626) about exact definitions of bit-security. We noticed that depending on how you account the adversary's work we have slightly (but noticeable) different security of Hash-based schemes. Be careful with your Lean-proofs and statements :)

Drove all the way to my local polling place and there was literally nowhere on my ballot to vote against a 24 team College Football Playoff and a 76 team NCAA Basketball Tournament. Had to write it all in

For the past four years whenever I’ve been in an interview loop and refused to do a take home assignment, it turns out that it was for cybersecurity reasons, not stubbornness. 😉

Recently, there seems to be the weird theory going around that hash functions are somehow more secure than lattices. There is no reason to believe that. It is perfectly possible that we live in a world where SHA and BLAKE are fully broken, but lattice assumptions are not.
This is not the consensus amongst cryptographers - sure, there might be some who think this, but to say that ECC is *likely* to break is definitely not the consensus. There is absolutely no evidence to support this statement. Likewise, to say that symmetric is more secure.



The most recent hack was

Hey, at least it wasn't EIP-3074

On my mission there were three times a year we just got together as a zone and hung out at a stake center until cerfew for reasons like this. Halloween: trick-or-treat, Fourth-of-July: Fireworks, New Years Day: Hangovers (totally not college bowl games)
IT'S GO TIME: Mormons Prepare For One Evening Of Year When Everyone Will Answer Their Door buff.ly/hPaazgw

AI Potter and the Prisoner of `us-east-1`.
we should have a famous data center like us-east-1 where we imprison all the bad AIs kinda like guantanamo