c3.17 Add agentic security review to delivery
As AI coding assistants and autonomous agents join the delivery process, your review must cover what they produce and the tools they can reach. Agent-generated code, dependencies, configuration, MCP servers, and prompts can introduce risks — leaked secrets, prompt injection, known CVEs — that classic scanners miss.
Implementation guidance
Review the agents and what they produce
AI coding assistants and autonomous agents now generate or modify code, dependencies, configuration, and even the Model Context Protocol (MCP) servers they call. Each can introduce risks classic scanners are not designed for: leaked secrets, prompt-injection paths, and known vulnerabilities. Add an agent-aware security review to your DevSecOps pipeline that evaluates both the agents you use and the artifacts they generate.
Use AWS Security Agent for design and code review
AWS Security Agent (part of AWS Continuum) builds context from your application's design, code, and requirements, then reviews it at multiple checkpoints — design review, automated code review on every pull request, and on-demand penetration testing — flagging verified security issues before they ship.
Practices to apply
- Treat agent-generated changes with the same scrutiny as human-written code: scan for vulnerabilities, secrets, and insecure configuration in the pipeline.
- Review the MCP servers and tools your agents can call, and apply least privilege to the credentials and permissions they hold.
- Test for prompt-injection and data-exfiltration paths in agent workflows, and fail the build on high or critical findings.
Related
This pairs with agentic penetration testing later in the lifecycle.
How to validate
- Do you review the security of the AI agents and assistants used in delivery?
- Are you scanning agent-generated code, dependencies, and configuration for vulnerabilities, secrets, and prompt injection before it ships?
- Do you assess the MCP servers and tools your agents can access and constrain their permissions?
Threats addressed
- AI assistants and autonomous agents can introduce vulnerabilities, leak secrets, or act on prompt-injection instructions that traditional scanners are not designed to catch.
- As code, dependencies, configuration, MCP servers, and prompts are increasingly agent-generated, the attack surface expands beyond hand-written code and the credentials the agents hold become high-value targets.