今天abc看了啥🤔
- 还 tm 上电视了…
- 大模型推理JWT说是(
- b23.tv/BV1cfS9BjEoz
科隆 Major 的积分窗口已经锁定,让我们掌声恭喜连复活赛(主办方给面子直邀的)都没打赢的 FaZe 建队以来首次无缘 Major! - ping0.cc被曝利用WebRTC静默上报用户真实IP
https://www.nodeseek.com/post-674661-1 - 今天是yau的生日。轉發這條消息自我的手機到5個群聊,yau就會獎勵你一塊華為手表,還會允許你打手槍。這個機遇不可特意去求!我試過了,是假的,還不會受到任何處分,還會被移出群聊,還會被記兩次大過。但今天真的是yau的生日。
- 螺老板没补的图我补上
-
- User-Agent
- 需要顺便说一句的是,源代码里的 TrustedBot 除了上文的 @nekonotificationbot(1190800416)外,还同时出现了 @WatchdogNextBot (6371744499),供参考。
- 为验证这一点,我们制作了一个PoC:一个LSPosed模块,将机器人ID和用户名替换为我们自己的信息,这样所有请求都会发送到我们的服务器上。通过这种方式,我们确认电话号码确实在被收集。每次登录都会如此。
该PoC可在此处获取: https://github.com/RomashkaTea/nekogram-proof-of-logging
https://t.me/EvolutionXOfficial/2488 -
- BREAKING: Nekogram is secretly transmitting your telegram account phone number to the developer
According to SOTA,
"The backdoor is hidden in the http://Extra.java file, which differs from the template uploaded to the repository. The obfuscated code sends data as an inline request to the @nekonotificationbot, leaving no trace. The same file implements account 'doxing' via several bots; it is possible that the leaked data is used to populate their databases."
Additionally, the creator of the Nekogram client, (presumably a Chinese national) was previously known for conducting DDoS attacks and unethical online behavior (including death threats against acquaintances).
Apparently, in the early versions of the client, de-anonymization was applied only to Chinese phone numbers, which could have been used for political surveillance;. However, it is now applied to all users.
Follow @TechLeaksZone -
-
-
- 什么 openclaw 都是小儿科,给我 instructkr/claude-code 擦皮鞋
-
-
- 哈哈哈哈,这简直是这 51.2 万行源码里最通人性也最“卑微”的代码
这段 negativePattern 暴露了 Anthropic 工程师在调教 Claude Code 时的三个核心策略:
1. “按闹分配”的止损逻辑:
当正则命中这些词(比如 wtf、piece of shit)时,系统会瞬间切换到一个 “高阶安抚模式”。它不再跟你争论逻辑,而是优先承认错误,并自动调用更强的 Opus 模型或开启更深的 CoT(思维链)来强行解决当前的 Bug,试图通过“超常表现”挽回用户的心。
2. AI 的“读心术”底层是正则:
大家觉得 Claude 懂人心,其实底层的“感知器”居然是这么一串脏话列表。这说明在极端的工程环境下,最原始的正则比复杂的情感分析模型更鲁棒、更省钱。
3. 防止模型“对骂”:
代码里还有一个配套的 SafeResponse 逻辑。一旦检测到用户抓狂,它会强制要求 AI 保持 “Extreme Professionalism(极端专业主义)”,绝对不允许 AI 回嘴。这就是为什么它有时候显得那么“温润如玉”——其实全是这几行正则给逼的!