One leaked key at a vendor turned into malware on developer machines. Cursor says an access key it issued to its inference provider, Fireworks AI, was exposed in a breach at Fireworks, and the attacker used it to start a chain that ended in the Cursor extension marketplace.
How the chain worked
The key opened a read-only storage bucket shared with the vendor. It held historical coding-agent traces from sessions where users had not turned on Privacy Mode. Most dated to 2025, and fewer than 0.01% of customers had unredacted API credentials in them. Those credentials gave the attacker access to several internal SaaS tools, including archived non-production code, limited support data and summaries of some agent conversations.
The trojanized extension
On Oct. 29, the attacker replaced Anthropic's Claude Code extension in Cursor's marketplace cache with builds 2.1.312 and 2.1.313. They stayed up for about 11 hours. The payload collected host details, tried to add an attacker SSH key and open a reverse tunnel, and installed persistence under names like com.apple.sysmond on macOS and SysMon on Windows. Cursor says up to 5,000 users were potentially exposed. Anthropic's own systems and the public Open VSX marketplace were not involved.
What to do
Uninstalling the extension is not enough because the persistence stays behind. Cursor published removal scripts for macOS and Windows, and affected users should rotate every credential reachable from those machines. Cursor says it rotated the exposed keys, rebuilt the marketplace and moved inference traffic away from Fireworks, and it has seen no malicious activity since Nov. 1.
Timing is everything
AI coding tools hold your repos, tokens and shell, so their vendors' vendors are now part of your attack surface. Turn on Privacy Mode, and never paste live secrets into an agent session.



