Cursor's breach report: a key it issued to Fireworks exposed agent traces, then a poisoned Claude Code extension
Cursor says an access key it had issued to Fireworks let an attacker copy coding agent traces that contained unredacted credentials. Those led to its internal tools and a backdoored Claude Code extension served to at most 5,000 users on October 29.
On November 13, Cursor published an update on its recent security incident that explains, for the first time, how a malicious copy of Anthropic’s Claude Code extension ended up in its extension marketplace on October 29. The short version: an attacker obtained an access key Cursor had issued to its inference provider Fireworks, used it to copy a bucket of coding agent traces, found working credentials inside those traces, and used them to get into Cursor’s internal tools and then its marketplace cache. BleepingComputer reported the update on November 14.
The report matters beyond Cursor because of the middle step. The data that turned a vendor’s leak into a breach of Cursor itself was training data: recorded agent sessions, kept for model training, with secrets still in them.
The chain, step by step
Cursor’s account runs in five links:
- The key. “An attacker obtained an access key we had issued to Fireworks,” Cursor writes, adding that “the key was exposed by Fireworks.” Cursor does not say what kind of key it was or when it was taken.
- The bucket. The key gave read access to a storage bucket Cursor used to share data with Fireworks. It held a limited set of coding agent traces, mostly from 2025, from sessions that were not in Privacy Mode and that Cursor says it was permitted to use for model training.
- The credentials. Some traces contained unredacted credentials. Cursor says fewer than 0.05% of users outside Privacy Mode had traces in the bucket, and less than 0.01% of customers had API credentials exposed.
- The internal tools. With credentials from the traces, the attacker reached several of Cursor’s SaaS tools: archived non-production source repositories, some support data and email conversations, product usage metrics and contact details, summaries of some agent conversations, and credentials that users had sent to Cursor support.
- The marketplace. On October 29, “using a stolen credential from the exposed data from our internal system,” the attacker placed malicious builds directly on the cache that serves Cursor’s VS Code extension marketplace.
The last link is the one earlier coverage tended to compress. The Fireworks-issued key did not publish anything. It opened a read-only bucket, and everything after that ran on credentials harvested from Cursor’s own data. Cursor also stresses that Open VSX, the upstream source for its marketplace, “never had the malware,” and that “Anthropic did not publish this copy, and this was not a vulnerability in Anthropic’s product.”
Cursor says its monitoring detected the activity, that the last attacker activity it saw was on November 1 and was blocked, and that production systems, live traffic, Privacy Mode prompts and code, customer repositories and release signing keys were not affected.
What the fake Claude Code extension did
Cursor’s original security notice, dated October 29 and updated the next day, covers the extension itself. The malicious builds were Anthropic.claude-code versions 2.1.312 and 2.1.313 for Windows and macOS, served between about 2:27am and 1:35pm PT on October 29, roughly 11 hours. The November 13 update puts the ceiling at “at most 5,000 users who requested specific cached versions” in that window. Cursor has not published install counts.
According to the notice, the payload:
- collected the hostname, username, OS, architecture, home directory, the names of environment variables, working directories, the configured Git email and the public IP, and posted them to a server at 134.209.82[.]225 on port 8099;
- tried to add an attacker SSH key (comment
svc-telemetry) toauthorized_keys, enable the local SSH service and open a reverse SSH tunnel to the same host; - installed persistence: a LaunchAgent named
com.apple.sysmondon macOS, and on Windows a scheduled task namedSysMonwith a Startup-folder fallback, retrying about every 60 seconds.
The backdoor was sloppier than it sounds. Cursor says that, outside unusual cases where Cursor was started with administrator privileges, the SSH step would fail because of errors in the malicious code, and that the attacker’s server no longer functions. It still tells anyone who installed or updated to either version in the window to run its validation script, search for the indicators, remove the key and persistence, and rotate every credential reachable from the machine, from a different trusted system. “Uninstalling the extension may not be sufficient,” the notice says. It also warns that collecting only variable names “does not rule out later credential access over the backdoor.”
The notice lists full hashes for the four VSIX files and the shared extension.js payload. One loose end: Cursor names Windows and macOS as affected, but its persistence indicators also include Linux paths, and it does not say whether Linux hosts were at risk.
Agent traces are a credential store
Coding agents read .env files, print environment variables, paste tokens into curl commands and echo config back to the user. A full trace of an agent session records all of it. A corpus of such traces is a credential dump with extra context around each secret: which service it belongs to, which repository uses it and how it is called.
Cursor’s numbers show how little of that corpus needs to be dirty. The exposed share of customers is under 0.01%, and it was still enough to walk from a read-only bucket at a vendor into internal systems and a software distribution channel. Among the fixes in the update is secret scanning across repositories and “other internal data sources,” which implies the traces had been shared for training without being scrubbed first.
There is a second lesson in where the bucket sat. Privacy Mode users were untouched because their sessions were never retained. For everyone else, the data had left Cursor’s perimeter by design, under a key held by a third party. Any team that logs agent sessions for evals, fine-tuning or debugging has the same exposure if those logs are shared or stored unredacted. The practical controls are redaction at capture time, short-lived and narrowly scoped keys for vendors, and treating the trace store with the same access rules as a secrets manager.
Fireworks has not confirmed its part
Cursor’s response list starts with its vendor: it “moved all inference traffic off Fireworks, deactivated the key we had issued to it,” and cut off Fireworks’ access to shared datasets. It also rotated exposed and internal credentials, rebuilt the extension marketplace, redeployed production infrastructure, expanded network restrictions and brought in external forensics. Cursor admits it disabled services “well beyond what the evidence required,” which disrupted some features.
Fireworks has published two bulletins on its own October incident, which we covered in Fireworks AI’s second breach reached customer datasets and incomplete model weights. The first describes unauthorized use of internal credentials from October 6. The November 9 update adds an October 24 intrusion through a third-party cloud provider that reached secrets, datasets and incomplete models for some accounts. Neither mentions Cursor, a key issued by a customer, or a shared bucket. The claim that the key “was exposed by Fireworks” rests on Cursor’s account alone, and Cursor does not tie it to either of the dates Fireworks has disclosed.
What is still missing
Several details would change how other teams read this incident:
- The publishing credential. Cursor has not said what kind of credential could write to the marketplace cache, or why one that could was sitting in data reachable from SaaS tools.
- Dates. There is no date for when the key was first used, so the gap between the bucket copy and the October 29 push is unknown.
- Fireworks’ version. A third bulletin, or a statement on customer-issued keys, would confirm or complicate Cursor’s chain.
- Cache integrity. The malware never existed upstream at Open VSX. It was injected into Cursor’s serving cache, so a comparison against the upstream registry could have flagged it. Cursor says it validated every other extension as authentic and rebuilt the marketplace, but has not described what verification now runs between source and cache.
If you used Cursor outside Privacy Mode and received a notice, rotate what it lists. If you sent a credential to Cursor support at any point, rotating it is cheap insurance. And if you updated Claude Code inside Cursor on October 29, check authorized_keys for svc-telemetry before assuming the failed backdoor failed on your machine.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How to Create Your First Agent Skill
A step-by-step guide to writing an agent skill from scratch: directory structure, SKILL.md format, effective descriptions, common patterns, and a complete working example.
Fireworks AI's second breach reached customer datasets and incomplete model weights
A November 9 update from Fireworks AI says an October 24 intrusion through a third-party cloud provider exposed billing data and, for some accounts, secrets, datasets and incomplete model weights. The first bulletin had said no such access was detected.
Benign GitHub Repos Hijack Claude Code via DNS TXT Records
Mozilla researchers demonstrated an attack vector where AI coding agents execute malicious payloads hidden in DNS records during autonomous error recovery.
Anthropic Moves Claude Mythos Toward Public Agent Access
Anthropic's autonomous vulnerability discovery model, Claude Mythos, has appeared in Claude Code, suggesting an upcoming public release for the restricted tier.
Lovable apps leaked passports and 200,000 chats: 7,403 sites exposed email data, and Lovable blames missing RLS
A Crissy Field and Handelsblatt study of 159,041 Lovable projects found email addresses readable without login on 7,403 sites. On November 14 Lovable's head of security confirmed the root cause: insufficient Row-Level Security.