Offense

Offense is a product of Offense Research, Inc. that gives coding agents security tools for auditing dependencies, inspecting source and GitHub Actions workflows, and testing proposed fixes. It returns structured evidence through a CLI, MCP and an HTTP API, and the agent uses its own model, repository access and permissions to act on it. The code is open source under Apache-2.0, with optional paid hosting and a self-hostable platform.

What It Is

Offense is a security toolkit for AI coding agents such as Codex, Claude Code, Pi, Hermes and Crush. It does not run its own agent or require a model-provider key. Instead it produces evidence such as advisory matches, scanner findings, coverage warnings and remediation prompts, and the agent investigates and edits code within the scope the user authorizes. The README states that a finding is not a confirmed exploit and that a partial scan is not a clean bill of health.

How the Workflow Runs

The documented flow runs from discovery to assessment, investigation, fix and verification. A local engine previews supported npm and pnpm lockfiles, then queries OSV and npm metadata. It writes scan.json and a remediation.md for the agent. The agent traces real usage, tests upgrades in isolation and rescans the changed inputs. A preview mode makes no network calls, and uploads, paid calls and code edits need separate approval. Saved hosted audits or job receipts can also be exported as a local PDF report.

Execution Modes

Three modes are documented. The local dependency engine needs Node.js 24+ and npm. The self-hosted platform needs Linux amd64, Node.js 24+, Git and Docker, and runs dependency audits plus Opengrep and Zizmor in isolated containers with Offense billing disabled. The hosted service offers managed execution through an Offense account. A starter GitHub Actions workflow can add pull-request dependency audits via an installation PR.

Agent Integration

A portable Agent Skill, scanning-with-offense, teaches compatible agents the preview, approval and verification workflow. The README lists project skill directories for agents including Pi, OMP, Hermes, Codex, Claude Code, Crush, Cursor, OpenCode, Copilot and Gemini CLI. A local stdio MCP server exposes offense.audit_dependencies, and a Codex plugin is built from the same skill and CLI. The CLI can be installed through Homebrew, a GitHub release archive, or npm-style commands.