decrypted · 15 november 2026 · vulnerabilities and patching · supply chain · ai and llm security

Cursor's leaked supplier key ended in a booby-trapped Claude Code extension

After this morning's stolen GitHub tokens, tonight's story is a cousin from AI tooling. Cursor, the AI coding editor, has published its account of how a key issued to one supplier ended with a booby-trapped copy of Anthropic's Claude Code extension reaching up to 5,000 developers. Anthropic's own product was not at fault.

A key, a bucket and a marketplace

Cursor said an attacker obtained an access key it had issued to Fireworks AI, an AI infrastructure supplier. The key opened a read-only storage bucket of coding-agent traces, the logged conversations between developers and their AI assistant. They were mostly from 2025, from sessions not in Privacy Mode, which Cursor was permitted to use for model training. A small number contained unredacted credentials.

Think of a hotel key cut for a contractor's cleaner and left in a skip. It opens a drawer, and the drawer holds more keys. One of those opened the loading bay where the post is sorted.

What the attacker did next

Those credentials reached some of Cursor's rented online software and archived, non-production source code. On 29 October, Cursor said, the attacker used a stolen credential to place malicious copies of the Claude Code extension in the cache that serves Cursor's extension marketplace. Versions 2.1.312 and 2.1.313 were live for about 11 hours. At most 5,000 users who requested those cached versions may have received them.

BleepingComputer reports that the extension tried to add an attacker's key to the machine's SSH settings, SSH being the encrypted way administrators log in to servers, and to switch SSH on. It also reports that researchers found coding errors that probably stopped it working on many machines. That is the reporter's assessment, not Cursor's. Cursor says the separate open-source marketplace never carried the malware, and the last attacker activity, on 1 November, was blocked.

The design decisions that mattered

Four choices turned a leaked key into a software supply chain incident:

Scope each supplier key to one job and give it a short life. Scrub secrets before a transcript is stored. Let a cache serve only what a signing step has vouched for.

What a UK organisation should do

Ask your developers whether anyone installed that extension on 29 October. Cursor warns that uninstalling alone may not be enough, so look for an unfamiliar key in the machine's authorised SSH keys and rotate anything the machine could reach. Then check your own estate: which AI vendors hold keys to your systems, which transcripts they keep, and whether Privacy Mode is enforced rather than optional. Opinion, not advice: a vendor's training data is also your exposure.

Also this week

Outlook flaw exploited before the fix. Microsoft patched CVE-2026-124194 on 10 November, the flaw's public reference number, and confirmed attackers were already using it. A crafted message file makes Outlook contact an attacker's server without the usual warning, and Windows may offer up the user's NTLM hash, a scrambled password that can be cracked offline or replayed to impersonate them. It scores 6.5 on the zero-to-ten CVSS severity scale and gives no code execution, but a stolen identity is plenty. The fix arrives in the November Office updates, not Windows Update. Microsoft has not said who is exploiting it. The design lesson is old: software should not volunteer credentials to any server that asks, so restrict outgoing NTLM and block file-sharing traffic to the internet.

A chipmaker's staff data, held by a cloud provider. TSMC told US regulators on 12 November that some current and former employee data had appeared on an external website without authorisation. It said it had found no evidence its own systems were compromised, and that the data matched records held in the system services of an unnamed cloud provider, unrelated to customer or operations data. Data on an external site is not proof of how it got there, so treat the cause as unconfirmed. The lesson for UK firms is plain: personnel records in a supplier's cloud are still your responsibility, so know which supplier holds them and who can reach them.

Ofcom's age-check enforcement reaches a provisional verdict. On 9 November Ofcom issued a provisional notice of contravention under the Online Safety Act to Sun Social Media, over pornographic content. It provisionally finds reasonable grounds to believe the firm failed to use highly effective age assurance, and failed to respond accurately to a statutory information request. The firm has 20 working days to make representations before a final decision. Ignoring the regulator's questions is treated as a separate failing. For anyone building age checks, the design test is proving age while collecting as little personal data as possible, since every verification database is a future breach.

Sources

If your suppliers hold keys to your systems and you are not sure what they can reach, get in touch.

More like this

Get the next post by email: subscribe to Decrypted. Double opt-in, unsubscribe any time, or take the RSS feed.

Prefer to listen? Decrypted on Apple Podcasts, or paste the podcast feed into any app.

Play here Loads Apple's player when you press it.