Cursor audits merged changes for security findings

Cursor’s team describes running a security review agent on every push to its main branch. The agent audits the changed code and sends high-risk findings to Slack. The authors report that this automation caught multiple vulnerabilities and critical bugs.

The timing matters. Running after merge gives the agent longer to investigate nuanced problems without making every pull request wait for its review. Each invocation starts a cloud sandbox and follows the configured instructions, models and MCP connections. A memory tool can preserve lessons between runs. Within the security workflow, the agent skips issues already discussed in the pull request before posting findings, reducing repeated reports.

This is one of several automations the team says it had been running on its own codebase for several weeks. The source also describes reviewer assignment, incident investigation and test coverage agents, but does not publish a particular vulnerability or a count of confirmed security findings for this reviewer.

The reported result is an operating review channel with useful catches. It does not establish complete security coverage, a false-positive rate or automatic remediation of every finding. Slack reports still give engineers information to inspect and act on.

Prompt
Suggested prompt (written by Agents In Use, not the author's)

Source: cursor.com · Jack Pertschuk, Jon Kaplan & Josh Ma