1
Why LLM Decisions Should Be Deterministic
LLM决策的确定性边界关键不是一致性，而是可审计性——每个决策都有可复现的实现。
TL;DR: I originally treated deterministic boundaries around LLMs as a consistency mechanism. I now think their real value is auditability. If the syst…
2
Show HN: Grepathy – Claude made a decision nobody approved
AI自作主张后不留痕迹？Grepathy将Claude的决策记录本地提炼为markdown commit，让AI行为可追溯
hey HN - Claude pre-created users in Clerk with null emails/names as "guest users" on a contract job. Wasn't in any plan. The CTO asked why, and I did…
3
CrashStealer 攻击披露：针对苹果 Mac 用户，瞄准 1Password 等 14 款密码管理器
输入邮箱或密码，一键查询是否在已知数据泄露中曝光，让Mac用户也能防范类似CrashStealer攻击
IT之家 7 月 14 日消息，网络安全公司 Jamf Threat Labs 最新报告披露 CrashStealer 漏洞，影响约 80 个加密货币钱包扩展和 14 款密码管理器。 苹果在收到 Jamf 通报后，已撤销相关恶意应用签名凭证。 Jamf 研究人员于 2026 年 5 月初在 Viru…
4
Audit a Cursor or v0-Built MVP Before You Launch
AI编码助手生成的代码可能暗藏三类独特漏洞，发布前务必深度审计。
AI coding assistants ship code that compiles, runs, and looks correct in review. They also ship three specific classes of vulnerabilities that aren't …
5
Automated Textbook Auditing with Multi-Agent LLM Systems
多智能体大模型系统自动审核教科书，AI赋能教育内容质量把控新范式
arXiv:2607.11276v1 Announce Type: cross Abstract: Ensuring the quality of educational materials requires more than standard proofreading: textbooks mu…
6
Towards Autonomous and Auditable Medical Imaging Model Development
医学影像模型开发的未来方向：如何在自主生成的同时保证全流程可审计？这篇论文提出了新框架。
arXiv:2607.10522v1 Announce Type: cross Abstract: Large language model (LLM) agents are beginning to automate machine learning engineering (MLE) by co…
7
Codex starts encrypting sub-agent prompts
OpenAI Codex 开始加密子代理提示，导致多智能体 V2 消息的审计追踪不可读，引发回归问题
Article URL: https://github.com/openai/codex/issues/28058 Comments URL: https://news.ycombinator.com/item?id=48905028 Points: 264 # Comments: 160
8
Toward Auditable AI Scientists: A Hypothesis Evolution Protocol for LLM Agents
为LLM智能体设计可审计的假设演化协议，让AI科学家过程透明可信
arXiv:2607.09195v1 Announce Type: new Abstract: Large language model (LLM) agents are increasingly expected to play a central role in AI-driven scient…
9
Your AI's logs can be edited. Its evidence shouldn't be
AI日志可被篡改，但证据应保持不可变，这篇文章深入探讨了AI系统证据架构的设计原则与挑战。
Article URL: https://github.com/thekiraproject/kira-project-site/blob/main/posts/2026-07-evidence-architecture.md Comments URL: https://news.ycombinat…
10
Show HN: Audit.md for UI/UX Audit
一款专为UI/UX审查设计的Markdown工具，让设计评审更高效清晰。
Article URL: https://www.typeui.sh/audit Comments URL: https://news.ycombinator.com/item?id=48880982 Points: 1 # Comments: 0
11
Show HN: AI Agent Audit for Free
免费开源的AI Agent安全审计工具，帮你扫描漏洞、守护代理安全，开发者必备。
Article URL: https://github.com/sinewaveai/agent-security-scanner-mcp Comments URL: https://news.ycombinator.com/item?id=48870646 Points: 1 # Comments…
12
Token-Flow Firewall: Semantic Runtime Auditing for Persistent AI Agents
提出Token流语义防火墙，为持久AI代理提供运行时审计，保障Agent安全。
arXiv:2607.08395v1 Announce Type: cross Abstract: Persistent AI agents extend large language models (LLMs) beyond single-turn interaction into long-li…
13
Alberta Ran 50 AI Agents in Parallel. Everyone Shared the Same Number.
50个AI代理并行协作，4.66亿行代码扫描从6年缩至20小时
Alberta published a case study last week about scanning 466 million lines of code across 1,280 applications using around 50 specialized Claude agents …
14
Building a Linux Security Audit Toolkit with Bash: From Firewall Checks to SSH Hardening
手把手教你用Bash打造Linux安全审计工具包，从防火墙检查到SSH加固，实操性强。
Building a Linux Security Audit Toolkit with Bash: A Practical Approach to System Hardening Linux is widely considered a secure operating system. Howe…
15
AI Security Audit Checklist: 15 Vulnerabilities Claude Found in Production Code
AI安全审计清单：Claude发现的15个生产代码漏洞，含攻击向量与严重等级，开发者必看
Most web applications contain at least one vulnerability from the OWASP Top 10. A typical security audit takes 2-3 weeks and costs upward of $10,000. …
16
FirstResearch: Auditable Question Formation for LLM Scientific Discovery Agents
提出可审计的问题形成方法，让LLM科学发现代理更可靠、更透明，研究新思路值得一看
arXiv:2607.05682v1 Announce Type: new Abstract: LLM systems for scientific discovery increasingly assist with ideation, literature synthesis, experime…
17
The Moving Target: A Longitudinal Audit of Trustworthiness Drift Across Twelve Checkpoints of Open-Source Chat LLMs
追踪开源聊天LLM十二个版本的信任度变化，揭示模型行为随时间漂移的潜在风险
arXiv:2607.02587v1 Announce Type: cross Abstract: Model cards quote trust-benchmark scores without recording when they were measured, and the same num…
18
Auditing the Audit: Five Failure Modes in Benchmark-Validity Audits
基准测试的审计本身也可能出错？这篇论文揭示了基准有效性审计的五种典型失败模式，对于构建更可靠的AI评估体系是关键警示。
arXiv:2607.02586v1 Announce Type: new Abstract: Governance frameworks ask AI providers and auditors for documented evaluation evidence, and perturbati…
19
Scan Any Public GitHub Repo for Security Issues in Seconds (No Account)
无需注册，秒级扫描任何公开GitHub仓库的安全漏洞，适用于遗留代码、外包项目或求职者作业审查。
TL;DR You can now paste any public GitHub repo URL and get a security grade back in seconds, no signup. It runs static analysis, secret detection, and…
20
Detecting Answer-Driven Reasoning in LLM-Based Educational Tutors via Truncated Chain-of-Thought Auditing
通过截断思维链审计，揭示LLM教育导师是否存在"答案驱动"的推理偏差，提升AI教学可信度。
arXiv:2607.04572v1 Announce Type: new Abstract: Large language model (LLM) tutors often produce fluent step-by-step explanations, but a correct and pe…