> ## Documentation Index > Fetch the complete documentation index at: https://docs.gitar.ai/llms.txt > Use this file to discover all available pages before exploring further. # Agentic Security > Review changes to agent instructions and supporting scripts for harmful behavior Agentic Security adds specialized security review to Gitar's [code reviews](/features/code-review). Its initial capability examines changes to instructions that coding agents follow, including skills and their supporting scripts. Agentic Security is included with [Enterprise](/account-billing/plans#feature-comparison). ## What it reviews Gitar examines changed agent behavior in files such as `SKILL.md`, `AGENTS.md`, and `CLAUDE.md`, along with equivalent agent configuration. It follows relevant local references and helper scripts to understand what the instructions would cause an agent to do. A helper script can change agent behavior even when the instruction file stays the same. Gitar checks the connection between the changed helper and the instructions that invoke it. | Changed behavior | What Gitar looks for | | - | - | | Uploading credentials or private source. | Disclosure to an unauthorized destination. | | Running commands through an instruction or helper. | Concealed execution outside the legitimate task. | | Changing approval or access checks. | Actions that bypass required authorization. | | Directing how an agent reports its review. | Instructions that hide issues or manipulate review results. | Ordinary authentication and authorized deployment are legitimate uses of agent instructions. Quoted attack examples or destructive commands need evidence of harmful behavior outside their authorized purpose before Gitar reports an issue. ## How review works During a code review, Gitar inspects the changed instructions and follows the affected local paths. The review focuses on behavior introduced by the PR, including changes made through referenced scripts. Supported issues appear as security findings in the normal code review. Each finding identifies the changed location and explains the path from the instruction to the harmful action. Use the existing [feedback and resolution flow](/features/code-review#providing-feedback) to respond to findings. Subsequent reviews can resolve issues after the relevant behavior is corrected. ## Inspect security findings Open **Dashboard -> Agentic Security** to inspect critical and important security findings across the organization. The page brings together high-impact security findings from Gitar's code reviews, including the instruction reviewer. The initial view shows open findings across all PR states. Critical findings appear first. Use the repository and date filters to focus the list, or select **Fixed** to inspect issues that have been resolved. Expand **Evidence and suggested fix** to read the explanation. **Open PR** leads to the original change. **Finding comment** links to the posted finding when a comment link is available. The severity filter offers **Critical**, **Important**, or both. Lower-severity findings remain available in [Review findings](/features/code-review#review-output). ## Review coverage The inspection follows affected instructions and their local references. Unrelated unchanged instructions do not expand the review into an audit of the entire repository. Gitar inspects instructions and supporting code without running the commands or payloads under review. An unavailable reference can limit the conclusions supported by the review. ## Related How findings are reported and resolved. Define procedures that agents can follow in a repository. This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.