Skip to content- Sources: primary, discussion
- Thread: Fireworks AI cloud provider breach and customer data exposure
- Summary: Cursor's 2026-11-13 incident update says an attacker used an access key Cursor had issued to Fireworks, exposed by Fireworks, to copy mostly 2025 coding agent traces from non-Privacy-Mode sessions, some holding unredacted credentials. With those credentials the attacker reached Cursor SaaS tools, archived non-production repositories and support data, and on 2026-10-29 placed malicious Anthropic.claude-code 2.1.312 and 2.1.313 builds on the cache serving Cursor's VS Code extension marketplace for about 11 hours. Cursor says fewer than 0.05% of users had traces exposed, fewer than 0.01% of customers had API credentials exposed, the last attacker activity was blocked on 2026-11-01, and it has moved all inference off Fireworks.
- Why it matters: Agent session logs kept for training can carry live secrets, so anyone who pasted keys into a non-Privacy-Mode Cursor session or installed those extension versions should rotate credentials and follow Cursor's remediation steps, since uninstalling alone may not be enough.
- Background: Cursor is a fork of VS Code and cannot use Microsoft's extension marketplace, so it serves extensions from Open VSX through its own cache, which is where the attacker wrote the tampered builds. Agent traces are recorded transcripts of coding agent sessions, including tool calls and command output, which is how pasted or printed secrets end up in them. Fireworks AI, an inference and training vendor, disclosed several breaches in October, and Notion earlier said transcripts it shared with Fireworks were exposed. BleepingComputer places the incident in the wave of attacks on AI developer platforms that Wiz calls Black October.
- Comments: BleepingComputer adds that researchers found coding errors that likely kept the SSH backdoor from working on many Windows and macOS machines where SSH was not already enabled.