Glossary
These terms are used throughout the playbook, from the first review tutorial to the sandbox reference and threat model.
| Term | Meaning |
|---|---|
| Host / guest | Your computer / the Linux virtual machine (VM) running the review. |
| Harness | The program that connects the model to files and commands, such as OpenCode, Claude Code, Codex CLI or Pi. |
| Workload | The harness and everything it starts, including dependency installers and generated code. |
| Target | The repository or application being reviewed. |
| Skill | An instruction pack the operator installs for the harness to use. It may include scripts and other support files. |
| MCP | Model Context Protocol, an interface through which an agent can use tool servers. Host-connected servers can grant access beyond the VM. |
| Reproducer | A small program used to check whether a reported vulnerability is real. The wrapper provides a separate VM for running it. |
| Witness | A failing test, a crash, or a request and its response that shows a flaw with harmless input on the real application. Validation needs a witness; see kinds of evidence. |
| Proof of vulnerability / proof of concept | A proof of vulnerability is a witness, as benchmarks such as ExploitGym use the term. “Proof of concept” is used loosely for both witnesses and exploits, so ask which is meant. |
| Exploit | A flaw turned into an attack that shows what an attacker gains. More than a fix decision needs, and the step model providers gate. |
| Egress / allowlist | Outbound network traffic / the destinations that traffic is permitted to reach. |
| Control plane | The settings and services that enforce permissions, network rules and VM lifetime. |
| Provisioning / clean template | Installing tools before project data arrives / a saved VM baseline used by reset. |
| Residual risk | A risk that remains after a control is applied. |