All seven modules
Module 06 · Nerv-CODE · Coding agents
Coding Agent Security
Your developers gave an AI agent write access to the repository. Treat it like the privileged insider it is.
Coding agents read issues, docs, dependencies and tool output, all of it untrusted, then write code and run commands with developer privileges. That turns prompt injection into code execution. Nerv-CODE monitors what the agent does and stops what it should never have been able to do.
What it covers
- Agent action log: read, write, execute
- Protected paths and blocked commands
- Indirect injection in issues, docs and PRs
- Secret access and network egress control
- Malicious content in deps and tool output
- Secrets caught before commit and before read
- Hallucinated and typosquatted packages
- Insecure patterns in agent-written code
- MCP server and tool inventory, scoped
- Authorship provenance and audit trail
Where it separates
- 01Prompt injection is a supply chain problem here.
- A malicious instruction buried in a dependency README or a public issue is read by the agent and executed with your developer's credentials.
- 02Provenance you can hand to an auditor.
- Which code was agent-authored, under whose authority, from which prompt, reviewed by whom. Increasingly the first question in a regulated code review.
- 03Guardrails at the action, not the prompt.
- Filtering the prompt is a losing game. We constrain what the agent is permitted to do, so a successful injection still cannot reach production.
- 04Tied into machine identity.
- The agent's CI/CD tokens and repository credentials are governed as non-human identities in Nerv-NHI, not as untracked developer secrets.
Nerv-CODE works alongside the coding agents and IDE assistants your development team already uses. Supported agents are confirmed at scoping.