Code review stops at the diff: CodeRabbit's security agent now traces dependencies across repos and tests exploits in a sandbox
Synopsis
CodeRabbit now runs GPT-6.1 Sol as the primary model in its security agent, and review reaches past a single code change to risks across files and repositories. Previously a pull request was examined on its own, while the shared libraries, dependencies and safeguards it touches sat elsewhere. Sol reads changes as developers work; in Deep Scan and CodeRabbit Secure it follows dependencies and component interactions, and CodeRabbit reports flagging security issues across thousands of repositories, with user-selected findings testable in an isolated sandbox and drafted into fix pull requests.
Interpretation
CodeRabbit's security review now looks past a single change to what that change moves elsewhere. Before, review faced one pull request at a time, while a change can touch shared libraries, dependencies and safeguards in other parts of an application. CodeRabbit reports that GPT-6.1 Sol has flagged security issues across thousands of repositories and organizations, and those findings enter the pull-request workflow developers already use.
The code that writes and the agent that reviews are not the same model. Review runs on GPT-6.1 Sol, which CodeRabbit uses as the primary model in its security agent, a frontier model built for reasoning that sits alongside established scanners rather than replacing them. Its reach comes in two layers: Deep Scan investigates dependencies and component interactions across a repository, and CodeRabbit Secure extends that to linked repositories such as a shared authentication library.
Before a fix appears, the system tries to show the problem can actually be exploited. For user-selected findings, that check runs in an isolated sandbox rather than against production systems. For supported issues it drafts a pull request with a proposed fix for a person to review and approve, and human reviewers still decide which changes ship.
Perspective
The setup is aimed at engineering and application-security teams that review pull requests, covering a single change, dependencies and component interactions across a repository, and linked repositories such as a shared authentication library. It is positioned alongside established scanners and runs independently of the agent that wrote the code.
The count of flagged issues comes from CodeRabbit's own account. Which cases count as supported issues, and how far sandbox verification reaches, are not spelled out in the summary; this report reads that summary as well.