Breakdown: keys pasted into old Cursor agent sessions led to a fake Claude Code extension
By Pritesh Ghodge, security engineer · Published 16 November 2026
Did one of your keys end up somewhere public?
The Abbi Labs quick check tests your live, public site as an anonymous visitor: secret-looking keys in your JavaScript, exposed .env and .git files, Supabase tables readable with your public key. It can't see your agent chats or your laptop, so do the machine checks below yourself. No sign-up.
No app handy?
No sign-up. Read-only public-resource and protocol metadata checks. No model prompts or tool execution. Scan only sites you own or are allowed to test.
Run on every deploy or use from Cursor / Claude Code via MCP →
Weekly AI-app security checklist
One short email a week: a practical check for apps built with Supabase, Next.js and AI tools (RLS, leaked keys, headers and more). Unsubscribe anytime.
Most of us have done it. The agent can't reach the database, so you paste the connection string into the chat. The deploy fails, so you drop the API key in and say "use this". It works, you move on, and the key now lives in a session log you never think about again. Cursor's incident report from 13 November 2026 shows how far that can travel. Old coding agent traces, some with credentials nobody had redacted, sat in a storage bucket shared with a vendor. When the vendor's access key leaked, those credentials became the attacker's way in, and it ended with a backdoored copy of the Claude Code extension in Cursor's marketplace.
Everything below comes from Cursor's own two posts (the 29 October security notice and the 13 November update) and BleepingComputer and The Hacker News coverage of the update. Where something isn't public, I say so.
What happened
- It started at a vendor. Cursor had issued an access key to Fireworks, an AI training and inference provider. Cursor says Fireworks exposed that key, and the attacker used it to copy a read-only storage bucket Cursor used to share data with Fireworks.
- The bucket held agent traces. They were coding agent traces, mostly from 2025, from sessions not in Privacy Mode, which Cursor was allowed to use for model training. A small number contained credentials that hadn't been redacted.
- Those credentials opened more doors. Starting from them, the attacker reached some of the SaaS tools Cursor uses. From there they got archived non-production source code repos, some customer support data and emails, product usage metrics and contact details, summaries of some agent conversations, and some credentials users had sent to Cursor's support team.
- Then a fake extension. On 29 October, using a credential from that exposed data, the attacker put tampered copies of Anthropic's Claude Code extension (versions
2.1.312and2.1.313) straight onto the cache that serves Cursor's VS Code extension marketplace. Anthropic didn't publish them, and Open VSX, the marketplace source, never had the malware. - What the extension did. Per Cursor's notice, it collected host details (hostname, username, OS, environment variable names, git email, public IP), added an attacker SSH key to
authorized_keys, tried to switch on SSH and open a reverse tunnel, and set up persistence (a LaunchAgent on macOS, a scheduled task on Windows). Turning SSH on needs admin rights, so on most Macs and Windows machines that step failed. The key, the persistence and the host report still went through. - How Cursor responded. It pulled the extension after about 11 hours, took the marketplace offline and rebuilt it, killed the vendor key, rotated exposed credentials (then all internal ones), moved inference off Fireworks, and published indicators of compromise and cleanup scripts. Cursor says the last attacker activity it saw was on 1 November, and it was blocked.
The numbers
All from Cursor's 13 November update unless marked:
- At most 5,000 users who requested the specific cached versions on 29 October may have received the malicious build.
- The fake extension was live for about 11 hours (2:27am to 1:35pm Pacific time on 29 October, per the notice).
- Fewer than 0.05% of users had coding agent traces exposed. All of them were outside Privacy Mode, and Cursor says it notified each one.
- Fewer than 0.01% of customers had API credentials exposed. Team admins got a list of the specific keys to rotate.
- This sits in a wider wave Wiz has called "Black October". BleepingComputer lists Railway, Northflank, Lightning AI, OpenCode and Warp, and says stolen Railway customer secrets were used for about $740,000 in fraudulent AI charges.
What's confirmed and what we don't know
- Not affected, per Cursor: prompts and code of Privacy Mode users, live traffic, production systems, the Cursor app and its official updates (no release signing keys were exposed), customer code repositories, other marketplace extensions, and plugins.
- Not public: how Fireworks exposed the key. BleepingComputer notes Fireworks disclosed several breaches in October, but Cursor's post doesn't say which one this was. Cursor also doesn't say which SaaS tools were reached.
- Anthropic's side. Anthropic said on Hacker News that this wasn't a compromise of its systems and only affected Cursor users. The versions on its official channels were clean.
How to check your own machine
Only check machines you own or are responsible for. If you installed or updated the Claude Code extension from Cursor's marketplace on macOS or Windows on 29 October 2026, do this now. If not, skip to the next section, which applies to everyone.
- Look for the attacker's key (Cursor's own check):
# macOS / Linux grep -qs svc-telemetry ~/.ssh/authorized_keys && echo "svc-telemetry key found" || echo "No key found" # Windows PowerShell $files = "$HOME\.ssh\authorized_keys", "C:\ProgramData\ssh\administrators_authorized_keys" | Where-Object { Test-Path $_ } if ($files -and (Select-String svc-telemetry $files)) { "svc-telemetry key found" } else { "No key found" } - Look for persistence. On macOS,
~/Library/LaunchAgents/com.apple.sysmond.plistand~/.config/.smt. On Windows, a scheduled task namedSysMonorsysmon.cmdin your Startup folder. The full list of indicators is in Cursor's notice. - If anything turns up, uninstalling isn't enough. The payload drops scripts that keep running on their own. Run Cursor's remediation script for your platform, consider rebuilding the machine, and rotate every credential that machine could reach, from a different, trusted computer.
Check the other end too
A key in an old agent session is one way out. A key shipped in your live site is another, and anyone can find that one. Paste your URL and Abbi Labs checks what a signed-out stranger can see: secret-looking keys in your JavaScript, exposed .env and .git, open Supabase tables. Read-only.
What to change, whatever tool you use
The extension was the scary part, but the root cause is ordinary: real secrets typed into agent chats and support tickets, kept for months, then copied somewhere you never see. Cursor isn't the only tool that stores sessions. Here's what I'd do this week.
- Rotate anything you ever pasted into an agent chat or a support ticket. Database URLs, Stripe keys, OpenAI or Anthropic keys, Supabase service_role keys. If you can't remember, assume yes for the ones that matter. For Supabase, follow the service_role key exposed guide.
- Check your data-sharing setting. Only sessions outside Privacy Mode were in the exposed bucket. In Cursor, check Privacy Mode is on (on team plans, ask your admin to enforce it). In other AI coding tools, look for the setting that lets them use your sessions for training, and turn it off for work that touches real credentials.
- Give the agent the name, not the value. Keep secrets in
.env(gitignored) or your secrets manager and tell the agent to useprocess.env.DATABASE_URLor$DATABASE_URL. If it needs to run something with a key, let the shell fill it in:# bad: the key is now in the chat log curl -H "Authorization: Bearer sk_live_51H..." https://api.stripe.com/v1/charges # better: the log only ever shows the variable name curl -H "Authorization: Bearer $STRIPE_SECRET_KEY" https://api.stripe.com/v1/charges - Use scoped, capped keys for agent work. A restricted Stripe key, a read-only database role, an AI provider key with a monthly spend limit. Where the provider supports it, lock keys to your IP or network, which is one of the fixes Cursor applied to its own service providers. A stolen key that only works from your server, with a $50 cap, is a much smaller problem.
- Treat extension updates like dependency updates. Install from the publisher's official listing, notice when an extension you rely on updates, and be suspicious of a coding agent extension that suddenly asks for more than it used to. Cursor's notice is a good reminder that a trusted name in a marketplace isn't proof of who built that version.
- Scan more than your repos. Cursor added secret scanning to its code repos and "other internal data sources". Do the same at your scale: run a secret scanner over your repo history, your
.cursor/and.claude/folders, exported chat logs and anything you paste into tickets. See secret keys in git history and API keys in mcp.json.
My take
The chain here is long: vendor key, training bucket, old traces, SaaS tools, marketplace cache, your laptop. But the first link that matters to you is short. Someone pasted a live credential into an agent session, and nobody redacted it. You can't control how a vendor stores its keys. You can control whether your agent chats contain anything worth stealing. Rotate what you've pasted, switch the agent to variable names, and put a spend cap on every AI key you own.
Sources
- Cursor, An update on our recent security incident, 13 November 2026 (primary; source of the incident chain and impact figures)
- Cursor, Security notice: malicious version of Anthropic.claude-code (2.1.312 and 2.1.313), 29 October 2026, updated 30 October (payload behaviour, checks, indicators, remediation scripts)
- BleepingComputer, Cursor confirms security breach led to malicious Claude Code extension, 14 November 2026 ("Black October" context, Railway figure)
- The Hacker News, Cursor Says Stolen Fireworks AI Key Led to Trojanized Claude Code Extension Incident, November 2026
- Hacker News discussion with Anthropic's statement, Backdoor in Claude Code for VS Code extension on the Cursor Marketplace