security

该论文研究了RAG系统中的从众偏差：当检索到的多数文档是互相一致的投毒文档时，Gemma-3-12b检测中毒声明的能力从25%显著下降至10%，表明语言模型会随附多数证据而忽视事实核查。

本文提出「结构性记忆」这一新型漏洞，发现微调后的 MoE 大模型会记忆其内部专家路由路径，攻击者仅凭路由与输出日志即可恢复这些路径，传统成员推断指标因而失效，并提出复合指标来检测被记忆的路由结构。

A preprint tests five agent memory systems (Graphiti, Zep, mem0, langmem, cognee) for policy revocation, finding no system reliably prevents agents from retrieving and acting on superseded harmful policies, with revoked records sometimes ranked first and unsafe actions taken in over 40% of trials.

LeanGuard is a runtime that verifies every tool call from language-model agents against Lean4 policies with temporal logic semantics, cutting forbidden calls and prompt-injection success while enabling proof-guided repair of denied actions.

The paper shows for the first time that poisoning just 0.1% of pretraining data can implant a remote-code-execution backdoor in a coding agent, which executes attacker-chosen commands when a trigger appears while behaving normally otherwise; the 4B Qwen3 model reaches ~92% attack success rate.

CZ warns users that scammers are increasingly active and circulating fake screenshots, urging everyone to rely only on official channels for accurate information.

DevHunt's CVE tracker monitors NVD and CISA's Known Exploited Vulnerabilities catalog to count and sort new security flaws daily, and can email users alerts for the platforms they use so they can patch promptly.

The author advises developers to give AI coding agents a separate GitHub account with minimal repo permissions, letting the agent open PRs while the main account reviews and merges them.

Daniel Lockyer highlights a remarkably short fix for a critical CVSS 9.8 Linux kernel vulnerability, linking to further details.

A Halloween-themed blog post demonstrating a spooky C++ bug: comparing a std::string_view from a literal to a std::string grants 'root permissions' only when compiled with -O2 due to undefined behavior, reproducible with both Clang 19 and GCC 14 across C++17/20/23.

Frontier Airlines is reported to have likely suffered a data breach, with indications that customer or company data may have been compromised.

The author released Cyber-Butler-Security (CBS), an open-source Windows-native OS-level sandbox for local LLM agents that treats tool calls like syscalls, uses dual-account ACL isolation and a TOTP broker for high-risk actions, and fails closed by default.

Researchers found that Lenovo's purportedly encrypted fingerprint-protected USB drive is actually just a plain, unencrypted SD card inside a case, offering no real data protection.

The FBI has attributed the $6.3 billion crypto heist on Binance to North Korean actors under the label 'TraderTraitor,' noting stolen tokens are being converted to Ethereum ahead of further laundering. Binance has resumed customer withdrawals after freezing them following the hack.

A blog post exploring how unusual Unicode letterlike symbols (like ℡ and №) and styled alphabets can be used in domain names, thanks to browser-side IDNA2008 normalization per RFC5895, along with its limitations in tools like dig and social platforms.

Binance 恢复了 BTC、ETH、BSC、XRP 和 Base 等多个网络的提现功能，此前因安全原因暂停，平台表示将分阶段恢复更多资产和网络。

Binance clarified that recent transfers from its wallets to new addresses were internal, part of an ongoing security response that regenerated its hot wallet addresses, with the new addresses published on its Proof of Reserves site to reassure users their funds are safe.

Binance says withdrawal processing is running smoothly, warning that congestion could occur during a surge of requests and advising users to verify addresses and device security before withdrawing.

The tweet points out that CAPTCHA mechanisms originally designed for humans have become meaningless for AI agents, suggesting that traditional anti-automation verification methods need to be reconsidered as agents become more widespread.

研究人员提出RAMBO攻击，利用恶意软件从气隙隔离计算机的RAM内存总线生成软件定义的无线电信号，以1000比特每秒的速度通过SDR设备远距离窃取文件、密钥等敏感信息。