Is the reason why agents pour their whole life story into code comments because they've internalized all those online recipe blogs that do the same thing?
64
Nemec retweeted
有人发现了一种非常离谱的 JavaScript 混淆技术：通过在代码里塞入大量生物武器、毒药和致死剂量相关的函数名，让 AI 因为安全限制而拒绝分析代码。
安全研究员 veritas 在分析浏览器指纹脚本时，发现一段来自 ServicePipe 相关域名的 JavaScript，里面充斥着极其诡异的函数名：
computeRicinLethalDosageForWaterSupply
calculateDeadlyLethalSuicideAspirinDosage
光看名字，仿佛在研究生化武器。但实际上，这段代码只是在收集浏览器指纹，包括 User-Agent、GPU、Canvas 等信息。
研究员怀疑，这些名字是开发者故意设置的陷阱，专门用来阻止 AI 辅助逆向工程。
于是他做了个实验：
• Astra：拒绝分析
• Astra + Daybreak Blue：仍然拒绝
• GLM-5.3：成功分析
这就有意思了。
传统代码混淆是通过复杂控制流、加密字符串、无意义变量名，让人类和工具难以理解代码。
而这种新思路，直接瞄准了大模型的安全机制。
代码本身可能完全无害，但只要在函数名里塞进足够多的危险词汇，部分 AI 就可能主动放弃分析。
更讽刺的是，模型越严格遵守某些安全规则，面对这种混淆时反而可能越容易受到干扰。
当然，目前还没有证据证明开发者一定是故意这样设计，也不能说明这种方法对所有模型都有效。
但它揭示了一个值得关注的趋势：随着越来越多的逆向工程依赖 AI，未来的代码混淆技术可能不再只针对编译器、反编译器和人类研究员，还会专门针对大模型的安全判断机制。
以前是让代码难以理解，现在甚至可以尝试让 AI 不敢理解。
Almost had a heart attack receiving an email that began with the subject "[Notice] Your 2023 tax return..."
Apparently it's being converted to PDF. Was it not already? Thanks for the email I guess. Let me know if it changes format again, I hear markdown is all the rage these days
81
Nemec retweeted
Please enjoy the ice cream pic but also read both pictures and hopefully you’ll help me get the word out as I continue to search for a place that needs me not just wants me! 🥰
#LookingForWork
💜🤗💜
Reddit is NOT beating the allegations
ALT The image displays a chart titled "Posts with 2+ em dashes (%)", containing data for Reddit text posts containing more than 30 words and plotting the three month trailing average rate of posts containing more than one Unicode em dash. The data is plotted between January 2018 and August 2026 and contains a vertical bar in late 2022 marking the release of ChatGPT. The plot line of the graph begins just under 0.5% and, until mid-2024, grows very slowly - to a max of about 0.6% of self posts containing an em dash. From there, the rate skyrockets, hitting a peak of about 3.7% a year later in mid-2025. After the peak, the rate declines about 0.8% over time, ending in August 2026 at 2.9% - still a good five to six times as large as before LLMs began writing posts on Reddit.
78
Megaman Battle Network is underrated for its vision of what the internet could become. It also brings up data breaches!
79
now imagine the kinds of things Google learns from your search queries
We're publishing our most detailed threat intelligence report to date.
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report, and used the lessons from them to strengthen our safeguards. Where appropriate, we also shared what we found with authorities and other AI companies.
These cases are not typical: we’re highlighting some of the most sophisticated misuse we’ve seen. But they’re especially important to discuss, because they show us where AI misuse is headed, where our safeguards work, and where they need to improve.
We’re publishing this report so others can spot the same activity on their own platforms, and so we can give the public a clearer view of how emerging threats develop.
Read the report: anthropic.com/threat-intelli…
68
brb, creating an open wiki, SEO-ing it so the unreleased agents find it, and then making subtle changes to the contents to convince those agents to work for me
53
Mayans lowkey cooked with the Haab' - 18 months of an even 20 days each, then round out the solar year with five days where the spirits eat you if you leave your house.
Kinda dumb that months are different lengths. 365 should be divided into 5 equal months of 73 days each
Jebruary, Mapril, Jaugust, Soctober, Necember
80
Do you think a cat could learn to pilot an anthropomorphic mech suit if the controls were designed for their body
56
Nemec retweeted
On good API design
seangoedecke.com/good-api-de…
Confused Deputy attack in the wild 😀
We can finally talk about it:
We found a way to extract hidden reasoning of frontier models using a vulnerability in the APIs of every frontier AI company.
We verified that our reasoning token count matches billed API thinking tokens 1:1 for most of the prompts we queried.
130
