A stolen key turned an AI coding tool into a backdoor

A malicious code extension spreading red tendrils into a laptop

Here is a small horror story for anyone who builds software with AI. In late October, attackers managed to place a backdoored version of a popular Claude Code extension into the marketplace of Cursor, one of the most widely used AI coding tools, where as many as 5,000 developers could have picked it up. They did not break Cursor’s core product. They did not find some genius exploit. They walked in on a borrowed key that should never have been lying around.

How it happened

Cursor’s own incident update, published on 13 November, traces the breach back to an access key it had issued to Fireworks AI, one of its inference suppliers. Fireworks exposed that key, and an attacker used it to read a shared storage bucket. That bucket held logs from coding-agent sessions that were not running in privacy mode, and those logs still contained authentication credentials that had never been cleared out. One careless key opened the door to a pile of other people’s credentials.

On 29 October the attacker used one of those stolen credentials to slip trojanised versions of a Claude Code extension, builds 2.1.312 and 2.1.313, into the cache that feeds Cursor’s marketplace. Anthropic’s genuine extension was never touched, and neither were the Open VSX or VS Code marketplaces. One security analysis reported that the malicious code gathered host details, planted an SSH key labelled svc-telemetry, and tried to open a reverse SSH tunnel back out, though that payload detail comes from a single outlet and is worth treating as provisional.

How bad, really

The scary number is 5,000 developers potentially exposed. Cursor’s own figures are smaller: trace exposure for fewer than 0.05 per cent of users, and API credential exposure for less than 0.01 per cent of customers. So 5,000 is best read as an upper bound, not a confirmed body count. To its credit, Cursor moved fast once it knew: it pulled the build, took the marketplace offline, had the attacker’s server taken down, killed the Fireworks key, shifted its inference traffic off Fireworks, rotated credentials and notified the people affected.

The real lesson: your AI tools are now your attack surface

This is the part worth sitting up for. The AI coding boom has quietly built an enormous new supply chain, and almost nobody is securing its seams. When you use a tool like Cursor, you are not trusting one company. You are trusting its inference suppliers, the storage where your session logs land, and every key passed between them. Each handoff is a new door.

And there is a nasty irony baked in. AI coding tools that log everything, in the name of making the assistant smarter, turn themselves into honeypots of credentials. The logs that leaked here were from sessions not running in privacy mode, which means the credentials were sitting there precisely because the tool was being helpful. Convenience and exposure were the same feature.

What to actually do

If you use Cursor, check whether extension build 2.1.312 or 2.1.313 is installed and remove it. Rotate any credentials you have ever shared with the tool or its support. Run AI coding sessions in privacy mode when you can, so your secrets are not sitting in a log bucket waiting for the next leaked key. And start treating your AI development stack the way you treat the rest of your infrastructure: as something with an attack surface that needs watching, not a magic box that only ever helps.

What this means

Nobody needed to defeat Cursor to poison it. They needed one supplier to fumble one key. That is the shape of risk in the AI tooling era: long, fast-growing supply chains, assembled at speed, held together by credentials that leak. This was a near miss with a quick cleanup. The next one may not announce itself as politely.

Edgar Friendly, calling the AI bullsh*t. More tools, more swearing: toptoolstack.com.

Did you know? The breach did not touch the AI model at all. It rode in on ordinary cloud plumbing: an access key, a storage bucket and some logs that were never cleaned. The most advanced tooling in software still falls to the oldest mistake in security.

Sources: BleepingComputer; Cybersecurity News; Red Side Security.

Get the free weekly stack: the AI tools and moves that matter, hype filtered out.Subscribe free →
Scroll to Top