Is Cursor safe: official documentation states that several guardrails are not security boundaries, with 6 categories of risk and a hardening checklist
Is Cursor safe? 6 risks, controls and best practices
Security vendor Adversa AI assesses that Cursor is safe for professional development, provided that controls that are off by default are turned on and hard boundaries are set outside the agent. On the data side, when Privacy Mode is enabled, Cursor states that customer data is not used for training, and it has zero data retention agreements with model providers; on the agent side, Workspace Trust is off by default, and since version 3.23 it can by default read files outside the workspace without approval, and the documentation itself acknowledges that several guardrails are only best-effort rather than security boundaries. The article outlines 6 categories of risk and states that the supply chain is a confirmed attack path, including an approximately 11-hour distribution of a Claude Code extension with an SSH backdoor on the official marketplace in October 2026; as of July 2026, Cursor had published 33 security advisories. The author recommends adjusting default settings and locking them with team policies, and, outside the agent, tightening credential permissions, denying outbound network access by default, reviewing agent configurations, and running unfamiliar code in one-time environments.