Tag: security
All the articles with the tag "security".
-
Anthropic, Same Week: Safety Report Backfires, De-identification Goes Full Auto
Anthropic pulled off two opposite moves in the same week: a safety report that backfired on customer conversation privacy, and a feature that let me push my de-identification tool to full auto-block.
-
Anthropic 同一週：安全報告翻車，去識別化工具升級完全體
Anthropic 同一週做了兩件相反的事：安全報告讓客戶對話隱私翻車，卻也讓我把去識別化工具升級成自動攔截的完全體。
- Updated:
Your Own Machine Is the Attack Surface: Leftover Credentials, Permission Flags, Untrusted Input
Everyone talks about AI security as whether the model will say the wrong thing. What actually bites you is the plaintext keys left on your disk, the permission flag you waved through, and the untrusted input you fed it.
- Updated:
你的電腦才是攻擊面：殘留憑證、權限旗標、不可信輸入
大家談 AI 資安都在談模型會不會說錯話，但真正會咬你的是本機殘留的明文金鑰、隨手開的權限旗標，還有你讓模型讀進來的不可信輸入。
-
PII Guard TW: A De-identification Tool Built for Taiwan
There's no off-the-shelf de-identification tool for Taiwan's PII formats. So I built one that keeps sensitive data on your machine while you send the rest to AI.
-
PII Guard TW：為台灣打造的個資去識別化工具
台灣的個資格式沒有現成的去識別化工具。所以我自己做了一個，讓機敏資料不離開你的電腦就能安全送 AI 處理。
-
Your AI Tools Are Now the Attack Vector: npm and Python Supply-Chain Backdoors, a $1000 Stolen-Key Bill, and a Scan-Before-You-Install SOP
Attackers are planting instructions in Claude Code, Cursor, and Gemini CLI configs so your own assistant runs the exfiltration script. Two supply-chain waves, one $1000 stolen-key bill, and the five minutes to spend before installing anything.
-
你的 AI 工具已經是攻擊媒介：npm/Python 供應鏈後門、外洩金鑰盜刷 1000 美元，與裝 MCP 前的掃描 SOP
攻擊者已經在 Claude Code、Cursor、Gemini CLI 的設定檔裡植入指令，讓你的 AI 助手幫他們跑竊取腳本。這篇串起兩波供應鏈攻擊、一次一千美元的盜刷，以及安裝前該做的那五分鐘。
-
Once It Ships, It's Gone: Three Times an Agent Misfired on an External System
Calendar invites, replies, and the internal notes on a hold event are all one-way doors. I got each of them wrong once this month, so here are the shapes of the accidents and the checks that catch them.
-
寄出去就收不回來：agent 動外部系統的三次誤操作
行事曆邀請、回信、佔位事件的內部備註，這三種動作都是寄出即無法收回。這個月我在這三處各犯一次，記下事故形狀與防呆判準。
-
Codex Built Its Own Evidence Package and Went to Argue With Google Support
A leaked Gemini backend key at PDT Learning got abused, no spending cap, and burned 1000 USD. I pointed Codex's browser automation at Google's live support to fight the charge, and it went so hard it built a 15-page evidence package and sent it over.
-
Codex 自己做了一份證據包，跑去跟 Google 真人客服吵架
PDT Learning 一支 Gemini backend key 外洩被盜刷、沒設 spending cap 怒噴 1000 USD，我用 Codex 的瀏覽器操作去跟 Google 真人客服爭費用，它認真到自己生出一份 15 頁證據包發給對方。
-
Opus 4.8 Cries 'Prompt Injection,' Codex GPT-5.5 Tracks Down the Real Cause: a Worktree Race
With multiple sessions open in one repo and no worktree, Opus 4.8 raised a false prompt injection alarm, told me to check for supply-chain attacks and rotate API keys, until Codex GPT-5.5 read the session log and pinned the real cause: just a worktree race.
-
Opus 4.8 搞烏龍報『提示注入攻擊』，Codex GPT-5.5 揪出 worktree race 真因
Opus 4.8 在我忘開 worktree 的情況下誤報「提示注入攻擊」，叫我去查供應鏈、rotate API key，最後 Codex GPT-5.5 查 session log 才揪出真因只是 worktree race。
-
From Miasma to Hades: How One Group Turned AI Tools Into a Supply-Chain Attack Vector
Two June 2026 npm/Python supply-chain attacks: Miasma backdoored Red Hat packages, then TeamPCP/UNC6780 upgraded to Hades, turning Claude Code, Cursor and 14 other AI tools into an attack vector. Includes self-check steps.
-
從 Miasma 到 Hades：同一個攻擊組織如何把 AI 工具當成供應鏈攻擊媒介
2026 年 6 月兩波 npm/Python 供應鏈攻擊：Miasma 入侵紅帽套件後門，TeamPCP/UNC6780 升級成 Hades，把 Claude Code、Cursor 等 14 款 AI 工具當成攻擊媒介。附自我檢查步驟。
-
Claude and Sensitive Data: Retention Windows, Enterprise Terms, and What to Check First
What to know before handling sensitive data with Claude: data retention windows, Enterprise contract protections, input minimization, coding agent controls, and alternative architectures.
-
用 Claude 處理敏感資料前必知
用 Claude 處理敏感資料前該知道的事：資料保留時間、Enterprise 合約保障、輸入最小化、Coding Agent 控制與替代架構。
-
Worried About Agents Going Rogue? Start Isolated, Hand Off Gradually
Someone asked what to do if you worry about AI agents going rogue. My answer: don't pick a highly autonomous lobster up front. Start in an isolated setting where you supervise and approve every step, then let go slowly.
-
擔心 agent 失控？先在隔離環境逐步放手
有人問擔心 AI agent 失控怎麼辦。我的回覆是：別一開始就選高度自主的龍蝦，先在你親自監督放行每一步的隔離環境裡操作，再慢慢放手。
-
Don't Brute-Force Sites That Need a Login—Your Token Is Your Account
Don't point an AI agent at sites that require a login. Your token is your account, and when the banhammer comes down you're gone. Pay a little for a third-party scraping API and offload the ban risk.
-
需要登入的網站別硬爬——你的 token 就是你的帳號
需要登入的網站，別叫 AI agent 去硬爬。你的 token 就是你的帳號，admin 一丟二向箔你就沒了。花點小錢買第三方爬蟲 API，把封號風險轉嫁出去。
-
Have Claude scan an MCP before you install it—it found 7 vulnerabilities and still said it's safe to install
This week AgentCrew Academy shipped a video about why you should run /security-scan before installing any MCP / npm / pip package. Tested a third-party MCP, Claude flagged 7 findings, then said "install is fine." This is the written companion to the video.
-
裝 MCP 之前先讓 Claude 幫你掃——找到 7 個漏洞，但說這個可以裝
講座當週我拍了一支影片，講「裝任何 MCP / npm / pip / clone 之前先跑 /security-scan」這件事。實測了一個三方 MCP，Claude 掃出 7 個漏洞，但綜合評估後說「可以裝」。這篇是影片的補充文字版。
-
AI Security Boundaries: The Lobster Craze, Presidio, and My Three Rules
When everyone is going wild with AI automation, I chose to understand my own capability boundaries first. Sharing Presidio for data privacy and my security practices with Claude Code.
-
AI 的安全邊界：龍蝦熱潮、Presidio、與我的三條鐵則
當所有人都在瘋 AI 自動化的時候，我選擇先搞清楚自己的能力邊界。分享 Presidio 隱私工具和用 Claude Code 工作的安全實踐。