Breach after breach: the trusted tooling layer is failing
The front page read like an incident log. A backdoor turned up in the Claude Code extension on the Cursor Marketplace (27 million installs, and the affected version has been yanked). Binance lost $6.3B in 20 minutes. Northflank and Scribd both disclosed breaches. The ChatGPT account takeover write-up showed anyone could be hijacked by sending them a link. Azure had five full cross-tenant compromises built on path traversal and IDOR bugs that commenters say have likely existed since the features launched. Molly, the Signal fork, hasn't shipped an update in 3 months and is vulnerable.
The pattern here: almost none of these are exotic. They are the boring bugs in distribution channels, shared tenancy and account recovery. The Claude Code case is the key one for builders. Extension marketplaces are now a delivery route straight into developer machines, and the buyer trusts the Anthropic name more than the store it ships through. Commenters were already asking how a vendor even communicates a fix like this across two marketplaces.
The Scribd thread shows the other side. Users complained that only some customers got emails, and others said the company has no sympathy from them. Slow or partial disclosure now costs trust as much as the breach does.
So what?
If you ship an extension, plugin or agent tool, assume it will be targeted and plan your revocation and notification path before you need it. If you run a platform, disclose to everyone at once and early. Customers are comparing you to every other breach they read today.
Read these
Backdoor in Claude Code for VS Code extension on the Cursor Marketplace
The $6.3B Binance hack was pulled off in just 20 minutes
Northflank Breached
Azure's Weakest Link – Five Full Cross-Tenant Compromises
Tell HN: Scribd Breached
We Took over ChatGPT Accounts
Molly(Signal fork) has not been updated since 3 months and is vulnerable