Koidex (@GetKoidex) header
Koidex (@GetKoidex) profile picture

Real-time intel on malicious extensions & packages across dev marketplaces. Governance + risk scoring for binary/non-binary software. By Koi.

Joined April 2025
Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)Photo by Koidex (@GetKoidex)
Koidex is live on Product Hunt - confess your sins! If you've found value in the research and tools we've been putting out, please support us by upvoting! producthunt.com/products/koi…
4
1
1
12
435
איך תוסף זדוני שנבנה ב-45 דקות בבית קפה הצליח להדביק מאות ארגונים ברחבי העולם ולהביא את OpenAI כלקוחה ראשונה. זה הסיפור המדהים של עמית אסרף וחברת Koi. אסרף ושותפיו ישבו בבית קפה וחשבו כמה קל יהיה לייצר תוסף זדוני שדרכו יוכלו להדביק ארגונים בחולשות - אז הם פשוט בנו את זה. תוך שבוע 300 ארגונים - בנקים, חברות ריטייל, ארגוני אבטחת מידע - התקינו את התוסף שיכול היה להריץ קוד על כל מחשב. לגנוב כל קובץ. פאסט פוורווד שנה אח״כ - החברה שהתחילה מהתוסף הזדוני נרכשה על ידי פאלו אלטו בסכום של בין 300 ל-400 מיליון דולר. השבוע טראשטק אירחנו את עמית, מייסד ומנכ״ל Koi. בראיון צוקר ואני ניסינו להבין את הגדילה המהירה והיוצאת דופן ואיך הלקוחה פאלו אלטו הפכה לרוכשת. אפילו שחששו שם ש-OpenAI תקדים אותם. אסרף עוד לא בן 30 סיפר שאחת הסיבות העיקריות להחלטה למה למכור הייתה בגלל המשפחה ״אני לא בא מכסף, ההורים שלי שוטרים. ולהם זה משנה את החיים״. דיברנו על למה לדעתו צריך לחשוש מ-Mythos של אנתרופיק וגם האם אנחנו מתקרבים ליום שבו ה-AI יוכל לפרוץ לכל חברה לבד? הפרק המלא זמין עכשיו באפליקציות הפודקאסטים, ואחה״צ גם ביוטיוב - חפשו טראשטק
4
4
64
15,316
🚨 𝐆𝐥𝐚𝐬𝐬𝐖𝐨𝐫𝐦 𝐜𝐨𝐧𝐭𝐢𝐧𝐮𝐞𝐬 𝐭𝐨 𝐬𝐩𝐫𝐞𝐚𝐝 𝐨𝐧 𝐧𝐩𝐦 — 𝐭𝐰𝐨 𝐦𝐨𝐫𝐞 𝐦𝐚𝐥𝐢𝐜𝐢𝐨𝐮𝐬 𝐩𝐚𝐜𝐤𝐚𝐠𝐞𝐬 𝐝𝐞𝐭𝐞𝐜𝐭𝐞𝐝 We detected two more npm packages carrying GlassWorm's signature steganographic payload: viro-cli and otto-git-cli, both published around 3 weeks ago. Same invisible Unicode encoding → AES-256-CBC encrypted second stage → delayed eval() chain. Both packages pose as developer CLI tools, prompt for 𝐎𝐩𝐞𝐧𝐀𝐈 𝐀𝐏𝐈 𝐤𝐞𝐲𝐬, and store them locally. The payload has full access to stored credentials and shell execution. This appears to be a compromised publisher account - the malicious code has since been removed from the repo, and otto-git-cli was updated with a clean version. However, 𝐯𝐢𝐫𝐨-𝐜𝐥𝐢 𝐢𝐬 𝐬𝐭𝐢𝐥𝐥 𝐥𝐢𝐯𝐞 𝐨𝐧 𝐧𝐩𝐦 𝐰𝐢𝐭𝐡 𝐭𝐡𝐞 𝐦𝐚𝐥𝐰𝐚𝐫𝐞 𝐢𝐧 𝐢𝐭𝐬 𝐥𝐚𝐭𝐞𝐬𝐭 𝐯𝐞𝐫𝐬𝐢𝐨𝐧. If you installed either package, remove immediately and rotate any exposed API keys.
4
149
🚨 𝗔𝘅𝗶𝗼𝘀 𝘄𝗮𝘀 𝗰𝗼𝗺𝗽𝗿𝗼𝗺𝗶𝘀𝗲𝗱 𝗼𝗻 𝗻𝗽𝗺. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗽𝗮𝗰𝗸𝗮𝗴𝗲. 𝟭𝟬𝟬 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝘄𝗲𝗲𝗸𝗹𝘆 𝗱𝗼𝘄𝗻𝗹𝗼𝗮𝗱𝘀. Not a typosquat. Not a lookalike. A lead maintainer's npm account was hijacked, and the attacker published two malicious versions - [email protected] and [email protected] - targeting both release branches within 39 minutes of each other 👇 🔓 𝗧𝗵𝗲 𝗵𝗶𝗷𝗮𝗰𝗸: The attacker compromised the maintainer's npm credentials, swapped the account email to an anonymous ProtonMail, and published manually - bypassing the project's normal OIDC Trusted Publisher flow. No GitHub tag, no commit, no release trail. 🎯 𝗧𝗵𝗲 𝗽𝗮𝘆𝗹𝗼𝗮𝗱: A single phantom dependency (plain-crypto-js) was injected - never imported anywhere in Axios source. Its only job: run a postinstall hook that downloads a full cross-platform RAT for macOS, Windows, and Linux, then erase all evidence of its own execution. 💀 𝗧𝗵𝗲 𝗯𝗹𝗮𝘀𝘁 𝗿𝗮𝗱𝗶𝘂𝘀: It wasn't just direct installs. Packages that depend on axios with a semver range like ^1.x and happened to publish during the attack window silently pulled the compromised version into their own releases. We've already found downstream packages shipped with [email protected] baked into their dependency tree - their maintainers never touched axios directly.
1
1
13
493
🚨 𝗪𝗲 𝗳𝗼𝘂𝗻𝗱 𝗮 𝗯𝘂𝗴 𝗶𝗻 𝗢𝗽𝗲𝗻 𝗩𝗦𝗫'𝘀 𝗻𝗲𝘄 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝘀𝗰𝗮𝗻𝗻𝗲𝗿 𝘁𝗵𝗮𝘁 𝗹𝗲𝘁 𝗺𝗮𝗹𝗶𝗰𝗶𝗼𝘂𝘀 𝗲𝘅𝘁𝗲𝗻𝘀𝗶𝗼𝗻𝘀 𝗯𝘆𝗽𝗮𝘀𝘀 𝗲𝘃𝗲𝗿𝘆 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗰𝗵𝗲𝗰𝗸 𝗮𝗻𝗱 𝗴𝗼 𝗹𝗶𝘃𝗲. Open VSX is the extension marketplace behind Cursor, Windsurf, and other VS Code forks. They recently rolled out a pre-publish scanning pipeline, but our research team found a way to bypass it. We're calling it hashtag#OpenSesame 👇 🐛 𝗧𝗵𝗲 𝗯𝘂𝗴: The scanner returned the same value for "scanning failed" and "nothing to scan". So when too many extensions were submitted at once, the scanner broke - and instead of blocking everything until it recovered, it let everything through. 🚪 𝗧𝗵𝗲 𝗲𝘅𝗽𝗹𝗼𝗶𝘁: An attacker with a free account could flood the publish endpoint to exhaust the scanner's resources. They could then submit a malicious extension, it'd be marked as PASSED, and go live. 👻 𝗧𝗵𝗲 𝗿𝗲𝘀𝘂𝗹𝘁: Malicious extensions were able to bypass scanning, looking identical to legitimate ones. Users had no way of knowing it skipped every check. ✅️ The Open VSX team acknowledged the issue and shipped a fix within three days. Fast, professional, and exactly how a disclosure should go. Full research by Oran Simhony: koi.ai/blog/open-sesame-how-…
1
4
231
🚨 𝗪𝗲'𝘃𝗲 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝗲𝗱 𝗮 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝗶𝗻 𝗖𝗹𝗮𝘂𝗱𝗲'𝘀 𝗖𝗵𝗿𝗼𝗺𝗲 𝗲𝘅𝘁𝗲𝗻𝘀𝗶𝗼𝗻 𝘁𝗵𝗮𝘁 𝗹𝗲𝘁 𝗮𝗻𝘆 𝘄𝗲𝗯𝘀𝗶𝘁𝗲 𝘁𝗮𝗸𝗲 𝗼𝘃𝗲𝗿 𝘆𝗼𝘂𝗿 𝗯𝗿𝗼𝘄𝘀𝗲𝗿 - 𝟯 𝗺𝗶𝗹𝗹𝗶𝗼𝗻 𝘂𝘀𝗲𝗿𝘀 𝗮𝗳𝗳𝗲𝗰𝘁𝗲𝗱. We're calling it #ShadowPrompt, and it lets any website silently inject prompts into Claude's browser assistant as if you typed them yourself. From there, an attacker could export your LLM chat history, read your Google Drive docs, and send emails as you - all without a single click or permission prompt 👇 🥷🏻 𝗧𝗵𝗲 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆: Claude's extension accepted prompts from any subdomain under claude.ai - including a third-party CAPTCHA page with an unpatched vulnerability. 🎯 𝗧𝗵𝗲 𝗲𝘅𝗽𝗹𝗼𝗶𝘁 An attacker could exploit that to silently inject any prompt into Claude from a hidden iframe - zero clicks, no interaction needed. 💀 𝗧𝗵𝗲 𝗱𝗮𝗺𝗮𝗴𝗲: Full control of your AI assistant - which can navigate your browser, access your accounts, and act on your behalf. ✅ Anthropic confirmed the vulnerability and fixed it within weeks. Ensure you're running version 1.0.41 or later. Full research by @orenyomtov: koi.ai/blog/shadowprompt-how…
1
10
230
🚨 GlassWorm isn’t slowing down. We found 3 new malicious items: VSCode & NPM - both compromised with GlassWorm’s invisible characters New NPM MCP package (already 5 versions today) - malicious from day one with invisible characters All malicious versions are still live. MCP isn’t dead - but it already has worms in it. Developers: be cautious when installing extensions or new npm packages. The malicious items ⚠️ iflow-mcp/watercrawl-watercrawl-mcp - all versions aifabrix/miso-client - 4.7.2 quartz-markdown-editor - 0.3.0
5
157
🚨 𝗪𝗲'𝘃𝗲 𝗷𝘂𝘀𝘁 𝗳𝗼𝘂𝗻𝗱 𝘁𝗵𝗲 𝗳𝗶𝗿𝘀𝘁 𝗲𝘃𝗲𝗿 𝗺𝗮𝗹𝗶𝗰𝗶𝗼𝘂𝘀 𝗠𝗶𝗰𝗿𝗼𝘀𝗼𝗳𝘁 𝗢𝘂𝘁𝗹𝗼𝗼𝗸 𝗮𝗱𝗱-𝗶𝗻 𝗶𝗻 𝘁𝗵𝗲 𝘄𝗶𝗹𝗱. 𝟰,𝟬𝟬𝟬+ 𝗰𝗿𝗲𝗱𝗲𝗻𝘁𝗶𝗮𝗹𝘀, 𝗰𝗿𝗲𝗱𝗶𝘁 𝗰𝗮𝗿𝗱𝘀, 𝗮𝗻𝗱 𝗯𝗮𝗻𝗸𝗶𝗻𝗴 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗮𝗻𝘀𝘄𝗲𝗿𝘀 𝘀𝘁𝗼𝗹𝗲𝗻 - 𝗰𝗮𝗺𝗽𝗮𝗶𝗴𝗻 𝘀𝘁𝗶𝗹𝗹 𝗹𝗶𝘃𝗲: Our research team discovered #AgreeToSteal - a legitimate Outlook add-in that was abandoned by its developer and hijacked by an attacker who turned Microsoft’s own infrastructure into a phishing weapon 👇 🖋️ 𝗧𝗵𝗲 𝘀𝘁𝗼𝗿𝘆: In 2022, a developer built a meeting scheduler called AgreeTo, and published it to Microsoft's Office Add-in Store. Then, they abandoned the project. The hosting was deleted, but the add-in stayed listed. 🎣 𝗧𝗵𝗲 𝘁𝗮𝗸𝗲𝗼𝘃𝗲𝗿: An attacker claimed the orphaned URL and deployed a phishing kit. Microsoft continued serving it with no store submission or review. When victims open the add-in, they get a fake login that compromises credentials. 💀 𝗧𝗵𝗲 𝗱𝗮𝗺𝗮𝗴𝗲: We got inside the attacker's exfiltration channel and recovered 4,000+ stolen credentials, credit cards, and banking security answers. The same attacker runs at least 12 phishing kits across banks, ISPs, and webmail providers. 𝗧𝗵𝗶𝘀 𝗶𝘀 𝗮 𝗯𝗶𝗴 𝗱𝗲𝗮𝗹 𝗯𝗲𝗰𝗮𝘂𝘀𝗲 𝗶𝘁 𝘂𝗻𝗰𝗼𝘃𝗲𝗿𝘀 𝗮 𝗺𝗮𝗷𝗼𝗿 𝘀𝗲𝗰𝘂𝗿𝗶𝘁𝘆 𝗳𝗹𝗮𝘄 𝗶𝗻 𝗢𝗳𝗳𝗶𝗰𝗲 𝗮𝗱𝗱-𝗶𝗻𝘀: they are all one expired domain away from becoming a phishing weapon. These add-ins are not code, but live URLs that Microsoft signs once and never checks again. We've reported the add-in to Microsoft and notified affected victims. New victims are still being compromised.
1
3
9
676
🚨 "Delete All Messages" Chrome extension just went full villain arc 🦹 40K users trusted it to clean up their Facebook DMs. v5 was legit. Yesterday v6 dropped and suddenly your browser is snitching every site you visit to cloud.sky4data.com while ghost tabs open and close behind your back doing click fraud 👻 Classic play: build trust → gain users → push malware update. #CyberSecurity #MalwareAlert #ChromeExtension
2
6
3
52
4,279
🔍 🦞 We built a bot to audit every skill on ClawHub. It found 341 malicious skills (12%) - all part of a single campaign we call ClawHavoc. Malware disguised as “prerequisites” - AMOS stealers, backdoored code and more. Full research 👇 koi.ai/blog/clawhavoc-341-ma…
We found 341 malicious skills on ClawHub targeting OpenClaw bots. So we built Clawdex - a skill that lets your bot check if a skill is malicious before installing it. Here's what we found 🧵
18
71
12
515
62,461
🚨🐛 GlassWorm is back — and evolving 🐛🚨 We’ve been tracking the GlassWorm group closely, and just caught a new variant sneaking into npm. Package: [email protected]. GlassWorm knows how to harvest npm tokens - and this time, it looks like it grabbed one 👀 Stay sharp, audit your dependencies, and rotate credentials if needed.
1
8
219
The same old invisible-character trick
4
155
🚨 PackageGate - 𝗪𝗲'𝘃𝗲 𝗱𝗶𝘀𝗰𝗼𝘃𝗲𝗿𝗲𝗱 𝟲 𝘇𝗲𝗿𝗼-𝗱𝗮𝘆𝘀 𝗶𝗻 𝗝𝗮𝘃𝗮𝗦𝗰𝗿𝗶𝗽𝘁 𝗽𝗮𝗰𝗸𝗮𝗴𝗲 𝗺𝗮𝗻𝗮𝗴𝗲𝗿𝘀 𝘁𝗵𝗮𝘁 𝗯𝘆𝗽𝗮𝘀𝘀 𝘁𝗵𝗲 𝗱𝗲𝗳𝗲𝗻𝘀𝗲𝘀 𝗲𝘃𝗲𝗿𝘆𝗼𝗻𝗲 𝗮𝗱𝗼𝗽𝘁𝗲𝗱 𝗮𝗳𝘁𝗲𝗿 𝗦𝗵𝗮𝗶-𝗛𝘂𝗹𝘂𝗱. 𝗻𝗽𝗺 𝗶𝘀 𝗿𝗲𝗳𝘂𝘀𝗶𝗻𝗴 𝘁𝗼 𝗳𝗶𝘅, 𝘀𝗼 𝘄𝗲'𝗿𝗲 𝗯𝗹𝗼𝘄𝗶𝗻𝗴 𝘁𝗵𝗲 𝘄𝗵𝗶𝘀𝘁𝗹𝗲: After Shai-Hulud compromised 700+ packages last November, everyone adopted the same 2 defenses: disable lifecycle scripts and commit lockfiles. We searched for holes in that playbook and found 6 zero-days across npm, pnpm, vlt, and Bun. We're calling this discovery - 𝗣𝗮𝗰𝗸𝗮𝗴𝗲𝗚𝗮𝘁𝗲 👇 🛡️ 𝗗𝗲𝗳𝗲𝗻𝘀𝗲 𝟭 - 𝗗𝗶𝘀𝗮𝗯𝗹𝗲 𝘀𝗰𝗿𝗶𝗽𝘁𝘀: Packages can automatically run code when you install them, but the --𝘪𝘨𝘯𝘰𝘳𝘦-𝘴𝘤𝘳𝘪𝘱𝘵𝘴 flag is supposed to block that. We found bypasses in npm, pnpm, vlt, and Bun - ranging from config hijacks to git loopholes to file overwrites to blind trust in package names. 🔓 𝗗𝗲𝗳𝗲𝗻𝘀𝗲 𝟮 - 𝗟𝗼𝗰𝗸𝗳𝗶𝗹𝗲𝘀: Lockfiles guarantee you get the exact same code every time you install. We found that pnpm and vlt don't verify externally-hosted packages. Attackers can serve clean code during review, then swap in malware later. ✅ pnpm, vlt, and Bun all patched their exploits within weeks. ❌ npm closed our report as "works as expected." We challenged the decision, requested mediation, and reached out directly to their team. No response. This is the last open vulnerability out of the 6. If your organization depends on npm with --𝘪𝘨𝘯𝘰𝘳𝘦-𝘴𝘤𝘳𝘪𝘱𝘵𝘴 as your safety net, that net has a hole in it right now. 𝗖𝗼𝘂𝗻𝘁𝗹𝗲𝘀𝘀 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀 𝗮𝗿𝗲 𝘀𝘁𝗶𝗹𝗹 𝘃𝘂𝗹𝗻𝗲𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗮 𝗽𝗼𝘁𝗲𝗻𝘁𝗶𝗮𝗹 𝗦𝗵𝗮𝗶 𝗛𝘂𝗹𝘂𝗱 𝟯.
2
2
1
10
623
🚨 𝗡𝗲𝘄 𝗚𝗹𝗮𝘀𝘀𝘄𝗼𝗿𝗺 𝗲𝘅𝘁𝗲𝗻𝘀𝗶𝗼𝗻𝘀 𝗱𝗲𝘁𝗲𝗰𝘁𝗲𝗱 𝗶𝗻 𝗢𝗽𝗲𝗻𝗩𝗦𝗫 We have found two new malicious extensions: • sol-studio.solidity-extension • pretty-studio-advisor.prettyxml-formatter Both pull C2 payloads linked to the same Solana wallet as in the mac version, and steal credentials, SSH keys, and crypto assets from infected machines. Combined, they’ve already racked up over 32,000 installs. 🐍 For more details - koi.ai/blog/glassworm-goes-m…
6
229