Cursor says key exposed at Fireworks led to malicious Claude Code extension
Cursor's incident report says attackers used stolen credentials to reach some SaaS tools and customer data, then planted a tampered extension up to 5,000 users may have received.
Key takeaways
- Cursor says an access key it had issued to Fireworks AI was exposed and used to copy coding agent traces, some containing unredacted credentials.
- With those credentials, the attacker reached some of Cursor's SaaS tools, archived source code, support data and credentials sent to Cursor's support team, the company says.
- A stolen credential let the attacker place tampered Claude Code extension versions v2.1.312 and v2.1.313 in Cursor's marketplace for about 11 hours on Oct. 29; up to 5,000 users may have received them.
- The malware tried to install an SSH backdoor; researchers say coding errors likely stopped it on many Windows and macOS machines where SSH was off, though it still sent machine details to the attacker.
- Cursor says Anthropic's product was not compromised, it believes the incident is contained, and it has moved inference off Fireworks.
Cursor, the AI code editor, says a tampered copy of Anthropic's Claude Code extension that turned up in its extension marketplace on Oct. 29 was the result of a wider breach that started with an access key it had issued to Fireworks AI, a provider it used for AI model training and inference. The company set out its findings in an incident report on its blog after what it called an extensive two-week investigation, and BleepingComputer reported on it on Friday.
Cursor says the vast majority of its customers were not affected and that it has contacted each affected customer directly. It believes the incident is contained: the last attacker activity it saw was on Nov. 1, and that attempt was blocked.
How did the attack unfold?
Cursor says Fireworks exposed the key. The attacker then used it to download data from a storage bucket, set to read-only, that Cursor shared with Fireworks. That bucket held records of coding agent sessions, most of them from 2025. They came from users who had not turned on Privacy Mode, and Cursor says it was allowed to use them for model training. Some traces contained credentials that had not been redacted. Cursor says fewer than 0.05% of users had traces exposed and fewer than 0.01% of customers had API credentials exposed.
Using those credentials, the attacker reached some of the software-as-a-service tools Cursor uses, the company says. From there they got at archived, non-production source code repositories, some containing business contact information, along with a limited amount of support data and email, usage figures and contact details, summaries of certain agent conversations, and credentials that some users had shared with Cursor support.
On Oct. 29, using a credential taken from that data, the attacker placed tampered copies of the Claude Code extension on the cache that serves Cursor's VS Code extension marketplace, according to the company. Cursor says at most 5,000 users who requested those cached versions that day may have received the malicious build, which was available for about 11 hours. The affected versions were v2.1.312 and v2.1.313, Cursor and SecurityWeek report.
What did the malicious extension do?
Researchers at Phorion and OpenSourceMalware found that a JavaScript loader had been added to the start of the extension's code, which otherwise was left unchanged, BleepingComputer reported in October. The loader ran when the extension started, without any action by the user. It collected details about the computer, added an SSH key controlled by the attacker, tried to switch on SSH access and open a reverse tunnel, and set itself up to keep running, according to BleepingComputer and SecurityWeek.
Phorion found coding errors that likely stopped the malware from turning on SSH on many Windows and macOS machines where it was not already running, BleepingComputer reported. Cursor gave a different reason, according to SecurityWeek: it said turning on SSH needs administrator rights that the extension lacks on most macOS and Windows systems, so that step failed and the attackers could not log in through the tunnel. Even where the backdoor failed, the malware still sent details about the machine to the attacker's server, according to BleepingComputer.
Was Anthropic's product compromised?
No, according to both companies. Cursor says Anthropic did not publish the tampered copy and that this was not a vulnerability in Anthropic's product. Open VSX, the source Cursor's marketplace draws on, never had the malware, Cursor says. Anthropic's Felix Rieseberg said the company had neither built nor published the tampered copy and that only Cursor users were affected, SecurityWeek reports. Anthropic also pushed a new version of its extension to avoid confusion, according to SecurityWeek.
What has Cursor done?
Cursor apologised to those affected and says it has moved all inference traffic off Fireworks and cut off Fireworks' access to the data it had shared, rotated exposed credentials and, as a precaution, all credentials across its internal systems, checked every other extension and rebuilt the marketplace, redeployed production infrastructure and brought in outside forensics experts. It says the Cursor app and its official updates, customer code repositories and the prompts and code of Privacy Mode users were not affected, and that no release signing keys were exposed.
Users who installed or updated to version v2.1.312 or v2.1.313 from Cursor's marketplace during the exposure window should follow the steps in Cursor's security notice, the company says, warning that uninstalling the extension alone may not be enough.
Fireworks itself disclosed several breaches in October, BleepingComputer reports, one in which API keys and tokens were exposed and another in which intruders got into developer machines by way of a hacked outside cloud provider. The note-taking company Notion has also said AI chat transcripts from a few workspaces were exposed through a Fireworks incident, according to BleepingComputer. The cited sources do not say which Fireworks incident, if any, involved Cursor's key, and did not include a response from Fireworks to Cursor's account.
Questions readers are asking
How did attackers get into Cursor?
Cursor says an access key it had issued to Fireworks AI was exposed by Fireworks. The attacker used it to copy coding agent traces, some of which contained credentials that led into some of the SaaS tools Cursor uses.
Which Claude Code extension versions were malicious?
Versions v2.1.312 and v2.1.313 served by Cursor's VS Code extension marketplace on Oct. 29. Anthropic's own channels were not affected, according to Cursor and Anthropic.
What should affected users do?
Cursor says anyone who installed or updated to those versions from its marketplace during the exposure window should follow its security notice, because uninstalling the extension alone may not be enough.
Were Privacy Mode users affected?
Cursor says prompts and code from users with Privacy Mode were not affected. The exposed traces came from sessions not in Privacy Mode.
Sources
- Cursor: An update on our recent security incident
- BleepingComputer: Cursor confirms security breach led to malicious Claude Code extension
- SecurityWeek: Malicious Claude Code Extension Served in Cursor Supply Chain Attack
- BleepingComputer: Malicious Claude Code extension on Cursor planted SSH backdoors
- #cursor
- #supply-chain
- #data-breach
- #developer-tools
- #claude-code
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Cybersecurity4 min
ConnectWise urges ScreenConnect client updates as SCKit targets technicians
Microsoft, Huntress and Blackpoint have analysed the toolkit used against MSPs. It is built to jump to technicians, but researchers say they have not seen that work.
Cybersecurity3 min
Landmark Group says Shukran loyalty members' personal data was accessed
Names, contact details and dates of birth were involved, the Dubai retailer says, but not card details, passwords or points. It has not said how many members were affected.
Cybersecurity3 min
Operation Endgame takes down Gootloader, CastleRAT and StealC servers
The Justice Department says Gootloader infected at least 25,000 computers and was used against at least 44 US ransomware victims. Europol counts 235 servers hit in all.
Cybersecurity2 min
TSMC says leaked employee data matches records at a cloud provider
The chipmaker says its own systems show no sign of compromise after a seller offered a 6.9GB file of employee records on hacker forums for three bitcoin.