SECURITY TOOLS FOR AGENTS

Your agent writes code.
Give it security tools.

Offense gives AI coding agents security tools to audit code, check dependencies, and build DevSecOps workflows. Connect your agent through the CLI, MCP, or API.

Open portal Set up your agent

Install the CLI

Homebrew

brew tap offense-research/offense https://github.com/offense-research/offense.git
brew install offense-research/offense/offense

macOS + Linux Includes Node.js

npx

npx --yes --allow-remote=all --package=https://offense.sh/downloads/offense-cli.tgz offense --help

Node.js 20+ Try it without a global install

npm

npm install --global --allow-remote=all https://offense.sh/downloads/offense-cli.tgz

Node.js 20+ Install globally, then run offense --help

Ubuntu

sudo apt-get update && sudo apt-get install -y curl ca-certificates
curl -fsSL https://deb.nodesource.com/setup_24.x | sudo -E bash -
sudo apt-get install -y nodejs
npm install --global --allow-remote=all --prefix "$HOME/.local" https://offense.sh/downloads/offense-cli.tgz
export PATH="$HOME/.local/bin:$PATH"
offense --help

Ubuntu 22.04 / 24.04 Installs Node.js 24 via NodeSource with sudo, then the CLI in your home directory. Review the setup script before running. Keep ~/.local/bin in your shell's PATH for future sessions.

Arch Linux

sudo pacman -Syu --needed nodejs npm
npm install --global --allow-remote=all --prefix "$HOME/.local" https://offense.sh/downloads/offense-cli.tgz
export PATH="$HOME/.local/bin:$PATH"
offense --help

Arch Linux Updates system packages and installs Node.js + npm, then the CLI in your home directory. Keep ~/.local/bin in your shell's PATH for future sessions. No AUR package required.

Hosted-service CLI Installation guide

13 hosted tools8 engines1 API key

Your agent. Connected to security tools.

offense / dependency auditRecorded fixture

Check my dependencies before I ship.

HIGHCVE-2022-25883

semver@7.5.1 — Regular expression denial of service

Proposed upgrade7.5.1 →to 7.5.2

Recorded public fixture. Source reachability and compatibility not verified.

WORKS WITH YOUR WORKFLOW

Claude CodeCodexCursorAny MCP client

01 / YOUR WORKFLOW

Keep your agent.
Give it a security toolkit.

A model can reason about your code. It also needs tools to check it. Offense connects agents to scanners and dependency intelligence, so security checks can be part of the work—not a separate handoff.

FOR THE AGENT YOU ALREADY USE

Audit before you ship.

Give your coding agent tools to check source code, dependencies, secrets, and workflows. Bring the findings back into the same conversation to investigate and plan a fix.

Connect your agent

FOR THE AGENT YOU WANT TO BUILD

Build your own DevSecOps.

Start with security tools instead of scanner integrations. Use Offense through the CLI or API to build a focused review agent around your repositories, instructions, and development process.

Read the integration guide

02 / IN PRACTICE

What can your agent check?

Give your agent a security task. It uses Offense to run the checks, then brings the findings back to your code for review and testing.

BEFORE YOU SHIP

Find vulnerable packages.
Plan the upgrade.

Your agent sends package names and installed versions to Offense. The audit matches them against known advisories and suggests upgrade targets.

YOU ASK YOUR AGENT
“Audit this project’s dependencies. Show me the advisories and which versions I should upgrade to.”

Your agent’s next step: Review whether the advisory affects your code, update the dependency, and run your tests. An upgrade suggestion is not a verified fix.

Offense audit / Recorded public fixture

High advisory match

semver@7.5.1

Regular expression denial of service

CVE-2022-25883

Input checked: npm package semver, version 7.5.1
Proposed upgrade: 7.5.1to7.5.2

Proposed upgrade. Source reachability and compatibility not verified.

BEFORE YOU COMMIT

Catch credentials
before they reach Git.

Your agent submits the source and config files you want checked. Offense runs Gitleaks to flag credential patterns without echoing secret values into the report.

YOU ASK YOUR AGENT
“Scan these files for API keys and tokens before I commit. Keep any secret values redacted.”

Your agent’s next step: Remove the credential from source and use a secret reference instead. If it was exposed, arrange revocation or rotation—not just deletion.

Gitleaks / Illustrative example

Credential pattern detected

A token in a config file.

Gitleaks flags a value that matches a credential pattern. It does not prove that the token is active.

Credential valueREDACTED

Input checked: Submitted source and configuration files
Suggested response: Remove from source. Rotate if exposed.

Pattern match for review. Submitted files checked; Git history uses a separate tool.

REVIEW A NEW ENDPOINT

Trace untrusted input
into a SQL query.

Your agent submits an endpoint’s source to Offense. Opengrep checks for risky code patterns, including supported within-file paths from request input to SQL construction.

YOU ASK YOUR AGENT
“Review this search endpoint for SQL injection. Explain the input path and propose a parameterized query.”

Your agent’s next step: Inspect the flagged query in context, replace string interpolation with bound parameters, and test the endpoint’s behavior.

Opengrep / Illustrative example

Potential SQL injection

User input becomes SQL.

The search value is interpolated into a SQL string instead of passed as a bound parameter.

Request inputSQL query

Input checked: Endpoint source code; the application is not executed
Suggested change: Bind values with a parameterized query.

Static signal, not a confirmed exploit. Review applicability and test changes.

BUILD A SECURITY AGENT

Review CI changes
with your own agent.

Build a review agent that submits changed GitHub Actions YAML to Offense. Zizmor checks the workflow for unsafe patterns, such as pull request text inserted into a shell command.

GIVE YOUR REVIEW AGENT A JOB
“Check this workflow for untrusted input in shell steps. Suggest a safer way to pass those values.”

Your agent’s next step: Propose a workflow change for review. You provide the trigger and review process; Offense supplies the security check.

Zizmor / Illustrative example

Untrusted input in shell

A PR title becomes shell code.

A contributor-controlled title is interpolated directly into a run step, where shell syntax can change the command.

PR titleShell step

Input checked: GitHub Actions workflow YAML
Suggested change: Pass the title through an environment variable and quote its expansion in the shell.

Static workflow check. Review the change; workflows are not executed.

0104

03 / BUILD WITH OFFENSE

Small agents.
Real security jobs.

Five things you can build. Your codebase, your rules, Offense under the hood.

01 / KEEP DEPENDENCIES CURRENT

Your dependency bot.

A Dependabot-style updater that follows your upgrade policy—not just the latest version.

Example job

“Audit our npm packages every Monday. Propose upgrades for vulnerable dependencies, run the tests, and open a PR. Ask before major upgrades; never auto-merge.”

02 / GIVE EVERY PR A SECURITY PASS

Your first reviewer.

Find risky code and exposed secrets before a human reviewer opens the diff.

Example job

“Scan changed source files on each PR. Investigate findings in context, then draft a review with the affected code and a suggested fix. Don’t approve or merge.”

03 / KEEP KEYS OUT OF GIT

Your secrets guard.

Catch credentials before commit. Flag what needs removal or rotation.

Example job

“Check staged file contents for credentials. Keep values redacted, pause for review, and explain how to remove a match or rotate an exposed key.”

04 / CHECK THE PIPELINE

Your CI reviewer.

Spot unsafe shell input and risky permissions in GitHub Actions.

Example job

“Review changed Actions YAML for risky permissions and untrusted shell input. Propose a safer configuration without running the workflow.”

05 / FIND THE EDGE CASES

Your API tester.

Probe a local API against its schema. Turn failures into test cases.

Example job

“Review our OpenAPI spec and run bounded tests against the prepared local API. Reproduce failures and propose regression tests. Never target production.”

Requires a prepared app, OpenAPI schema, Docker, and scanner image.

You build the agent. We supply the checks. These are starting points, not prebuilt bots. Bring your runtime, repository access, triggers, tests, and approval rules.

04 / THE TOOLKIT

Good tools.
One way in.

Established security engines, ready for your agents.
One API key. Prepaid tool calls.

In Offense

Trivy

Dependency vulnerabilities

In Offense

Opengrep

Source code security

In Offense

Zizmor

GitHub Actions security

Browse all 13 toolsExplore the catalog

Source code security

Opengrep / Source files

In Offense

Dependency vulnerabilities

Trivy / Lockfiles

In Offense

Exposed secrets

Gitleaks / Source files

In Offense

Infrastructure configuration

Trivy / IaC files

In Offense

SBOM vulnerabilities

Trivy / SBOM

In Offense

Container image vulnerabilities

Trivy / Image archive

In Offense

Git history secrets

Gitleaks / Git bundle

In Offense

npm package risk

GuardDog / npm archive

In Offense

PyPI package risk

GuardDog / PyPI archive

In Offense

OpenAPI security policy

Spectral / OpenAPI

In Offense

Container image provenance

Cosign / Signature bundle

In Offense

Passive web checks

ZAP / HAR capture

In Offense

GitHub Actions security

Zizmor / Workflow YAML

In Offense

Scanner catalog by Offense. Open portal

Local tools

Via the Offense skill

Source + workflows

Opengrep + Zizmor

API testing

Schemathesis / Docker

05 / OUR DIRECTION

Security should move
at the speed of agents.

Agents are changing how software gets built. We think security should be part of that workflow from the start.

We’re building toward a security platform and harness for agents: a place to equip them, build security workflows, and make those workflows your own.

Today, we’re starting with the tools. Code audits, dependency checks, and the building blocks for your own DevSecOps agents. That’s the first step, not the finish line.

Building in this direction? Talk to us

06 / GET STARTED

Your next security check
starts with a prompt.

Copy this into your coding agent.

Set up Offense from https://offense.sh/llms.txt

Prefer the terminal? Install the CLI

Open portal Read the docs

07 / COMMON QUESTIONS

Before you connect.

The tools, the setup,
and where your agent takes over.

What is Offense?

Offense is a security toolkit for AI coding agents, built by Offense Research. It provides dependency audits, source and workflow checks, and other security tools through CLI, MCP, and HTTP API. Your agent investigates the findings, makes authorized changes, and runs your tests. The broader agent-security platform and harness is the direction we’re building toward.

Which coding agents can use Offense?

Offense works with coding agents such as Claude Code, Codex, Cursor, Pi, OMP, and Hermes through skills, CLI commands, or MCP. Your agent keeps its existing model, repository access, and permissions. Follow the per-agent setup instructions; a chat-only client without tool access cannot run the CLI.

Can I build a dependency bot or PR reviewer?

Yes. Use Offense’s tools to build a dependency-upgrade agent, a security-focused PR reviewer, a secrets guard, a CI reviewer, or an API tester. These are workflows you build, not preinstalled bots or drop-in replacements for Dependabot or human reviewers. You supply triggers, repository access, instructions, tests, and approval rules. Start with the documented agent recipes.

Is Offense free and open source?

The Offense source is public, with an Apache-2.0 root license and component notices. Basic npm/pnpm dependency advisory audits are free and need no Offense account or model key. Optional hosted tools use quoted, prepaid per-tool pricing. Check the current catalog for availability and prices; approve a spending ceiling before a paid call.

Does Offense upload my source code?

It depends on the workflow. Local dependency analysis queries public advisory and package metadata without uploading source code. Hosted source checks require selected files or artifacts to be sent to the service. Preview and approve that scope first. You can also self-host Offense. Installing the CLI or skill does not authorize uploads, active testing, or paid scans.

How do I install Offense?

Use the Homebrew, npx, or npm commands above for the hosted-service CLI. Homebrew supports macOS and Linux and supplies Node.js; npm and npx require Node.js 20+. These commands install or run the client, not scanner engines or your agent’s skill. To connect a coding agent, use the agent setup prompt and start a new session afterward.

Your agent writes code.Give it security tools.

Install the CLI

Homebrew

npx

npm

Ubuntu

Arch Linux

Keep your agent.Give it a security toolkit.

Audit before you ship.

Build your own DevSecOps.

What can your agent check?

Find vulnerable packages.Plan the upgrade.

semver@7.5.1

Catch credentialsbefore they reach Git.

A token in a config file.

Trace untrusted inputinto a SQL query.

User input becomes SQL.

Review CI changeswith your own agent.

A PR title becomes shell code.

Small agents.Real security jobs.

Your dependency bot.

Your first reviewer.

Your secrets guard.

Your CI reviewer.

Your API tester.

Good tools.One way in.

Dependency vulnerabilities

Source code security

GitHub Actions security

Source code security

Dependency vulnerabilities

Exposed secrets

Infrastructure configuration

SBOM vulnerabilities

Container image vulnerabilities

Git history secrets

npm package risk

PyPI package risk

OpenAPI security policy

Container image provenance

Passive web checks

GitHub Actions security

Local tools

Source + workflows

API testing

Security should moveat the speed of agents.

Your next security checkstarts with a prompt.

Before you connect.

Your agent writes code.
Give it security tools.

Keep your agent.
Give it a security toolkit.

Find vulnerable packages.
Plan the upgrade.

Catch credentials
before they reach Git.

Trace untrusted input
into a SQL query.

Review CI changes
with your own agent.

Small agents.
Real security jobs.

Good tools.
One way in.

Security should move
at the speed of agents.

Your next security check
starts with a prompt.