On September 23, Cursor released two new development bots: Rollouts, which watches a change all the way from merge to production, flags regressions, and tries to restore a healthy state; and Security Reviewer, which runs a security pass on every pull request and produces vulnerability explanations and fix suggestions. Both are aimed at Teams and Enterprise plans and are turned on from Cursor's automations tab.
Rollouts: writing the monitoring plan before the merge lands
According to Cursor, Rollouts "watches a change from PR to production, flags regressions, and acts to restore a healthy state" — in other words, it tracks a change from pull request through to deployment, flags regressions, and pulls the system back to a healthy state.
It connects to three kinds of systems: code hosting, deployment pipelines, and telemetry tools such as Datadog, Grafana, and Honeycomb. The process runs in two stages. Before the merge, it reads the code diff, assesses the risk, and generates a monitoring plan from that; if the metrics touched by the change have no instrumentation at all, it warns before the merge happens. After deployment, it compares live data against a baseline, can spot regressions that show up only on a single endpoint or in a single region, and tries to separate "expected changes" from actual problems — so that, say, adjusting a rate-limit threshold isn't mistaken for an incident.
Cursor says feature-flag integration and awareness of release schedules are still on the roadmap. Exactly what "restoring a healthy state" involves — how far an automatic rollback goes, and whether it requires human confirmation — is left vague in the announcement, so that will have to wait for further documentation.
Security Reviewer: tracing the code's data flow
Security Reviewer is described as something that "runs on every PR, reads the change in the context of the whole codebase." It traces where data enters and exits along the code flow, which Cursor says sets it apart from scanners that rely purely on rule matching. Its detection scope covers:
- Injection flaws such as SQL, command, template, and LDAP injection;
- Authentication and authorization defects;
- Credential leaks;
- Insecure deserialization and redirects;
- Dependency vulnerabilities and infrastructure configuration issues.
Cursor's own figures show the average review time dropping from 4.8 minutes to 3.8 minutes, and the comment acceptance rate rising from 45%-50% to 60%-70%. The announcement doesn't say what baseline those numbers are measured against — an earlier version of its own tooling, or a competing product — nor does it disclose sample size or the measurement period. An "acceptance rate" only means developers adopted the bot's comments; it isn't the same as a vulnerability-detection rate, and the two shouldn't be conflated.
Where it sits among similar tools
The AI code-review field has gotten crowded this year. Anthropic added a /security-review command to Claude Code in 2025, along with a matching GitHub Action that runs an automatic security check on every PR. GitHub Security Lab just published an LLM-driven fuzzing pipeline this week that runs the whole way from finding entry points to writing up vulnerability reports. Cursor's own Bugbot also got an update back in June, claiming more than 3x faster speed, a 22% cost cut, and 10% more bugs found.
The entry points differ. Claude Code's security review and Bugbot both stop at "before merge"; GitHub's pipeline leans toward security research, aimed at digging deep into mature open-source projects. Cursor has taken this a step further, extending Rollouts into the post-deployment phase — territory that belongs to SRE and operations teams. That space has long been occupied by things like Datadog's Watchdog and the anomaly detection built into various APM tools; Cursor's difference is that it has already read the code behind this particular change, so it knows which metrics to watch.
For teams, flipping the switch on is only the first step. Rollouts needs access to deployment systems and monitoring platforms, and "restoring a healthy state" could, in practice, mean directly triggering a rollback. Whether a large company's change-management process can accept a bot holding that kind of access is a harder problem to solve than the feature itself. Cursor has not yet published details of its permission model or audit-log design.
Sources: CocoLoop, Cursor's official blog, the GitHub blog, Anthropic's Claude Code documentation; verified against Security Reviewer's review time and comment acceptance rate, plan availability, and detection scope.