Security Scanning is a skill for reviewing the security of code, Git changes, and agent skills with Codex or Claude Code. It combines search rules with contextual analysis by the agent. Reports include findings, evidence, suggested fixes, and the actual coverage of the review.
Review an entire repository, staged changes, branch changes, or a supplied package. 18 areas of skill and AI security, including obfuscation, are complemented by 8 application code checks.
This repository contains one complete skill package. The included Python scripts prepare the input and validate reports; using the skill requires an agent and the tools listed in the requirements.
Repository structure
security-scanning-skill/
├── plugin.json
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── .agents/plugins/marketplace.json
├── skills/security-scanning/
│ ├── SKILL.md
│ ├── agents/openai.yaml
│ ├── references/
│ └── scripts/
├── scripts/install_skill.py
├── tests/
├── docs/
├── LICENSE
└── README.md
One shared skill for Codex and Claude Code. The manifests describe the same package, and both marketplace catalogs point to its root directory. Codex uses the portable plugin.json; Claude Code uses .claude-plugin/plugin.json. The root LICENSE covers the entire package.
Installation
After obtaining a license, choose either plugin installation from GitHub or standalone skill installation from a downloaded package. One method per agent is sufficient; using both may make two copies available.
Codex plugin
codex plugin marketplace add https://github.com/rkolcz/security-scanning-skill.git
codex plugin add security-scanning@rkolcz-security
Claude Code plugin
claude plugin marketplace add https://github.com/rkolcz/security-scanning-skill.git
claude plugin install security-scanning@rkolcz-security
These commands register the rkolcz-security marketplace from this repository. You can also register a downloaded package locally: open a terminal in its root directory and use . instead of the GitHub URL. For a local installation, keep the downloaded directory so that the source remains available. Download the entire package, including the hidden manifest directories.
Standalone skill
Download the entire package and open a terminal in its root directory, next to this README.md. The installer uses only the Python standard library. It copies the shared skill and LICENSE into the selected agent's skill directory.
Codex:
python3 scripts/install_skill.py --host codex
Claude Code:
python3 scripts/install_skill.py --host claude-code
The default installation paths are ~/.agents/skills/security-scanning/ and ~/.claude/skills/security-scanning/. SKILL.md sits directly inside that directory. To install the skill only in a specific project, add --project followed by the path to an existing project. Use --skills-dir to select a custom skill directory.
You can also copy skills/security-scanning/ into the agent's skill directory manually and include the root LICENSE in that copy. The instructions and scripts work independently of the plugin manifests.
Updates
For a standalone skill, run the installer from the new package with the same arguments. It replaces only security-scanning/, removes outdated files from the active installation, and keeps the previous version in skill-backups/ next to the skills/ directory. It prints the backup path. Other skills remain unchanged.
The installer records the update state before moving the active copy and prevents simultaneous installations into the same location. If the process is interrupted, run it again with the same arguments: it detects the unfinished operation and restores the previous installation or retains the new copy if it was already put in place. The --recover-only option performs recovery without installing an update. The skill is briefly unavailable between the two directory moves. If the process is killed at that point, it remains unavailable until the installer is run again and recovery completes. This mechanism does not guarantee protection against disk failure or power loss.
A hidden .security-scanning-installer/ directory containing a lock file remains next to skills/. During an operation, it also holds the recovery state and the copy being prepared. This is the installer's working directory, not a second skill.
A symbolic link used for the selected parent skill directory is resolved to its real path. A symbolic link in place of the target security-scanning/, in the installer state, or in the source package causes the installer to refuse the operation. The installer does not replace a directory belonging to a skill with a different name.
Update plugins through the agent's plugin manager. This release is 2.3.0 in all manifests, skill metadata, and the license. Future releases should use new version numbers; historical reports retain the versions used to produce them.
When switching from a standalone skill to a plugin, move the previous installation out of skills/ to a backup location outside the agent's discovery directories. This prevents the old skill from remaining available alongside the plugin. Global and project installations are separate locations; check both.
After installing or updating, start a new agent session in the project you want to review.
Your first scan
Open your project in the agent and invoke the skill.
Codex
$security-scanning Review the security of the current repository. Include evidence, suggested fixes, and any areas that were not covered.
Claude Code — standalone skill
/security-scanning Review the security of the current repository. Include evidence, suggested fixes, and any areas that were not covered.
Claude Code — plugin
/security-scanning:security-scanning Review the security of the current repository. Include evidence, suggested fixes, and any areas that were not covered.
In Codex, you can also select the installed plugin from the plugin list and ask the agent to use its skill.
The agent prepares a copy of the selected code, searches for suspicious content, and evaluates it in the context of the project. You receive a readable report. You can also request JSON output or ask the agent to save the report to a file.
Choose a scope
| When to use it | What the review covers | Example request |
|---|---|---|
| Getting to know a project | Current repository files; in Git mode, tracked files and new files that are not ignored. | “Review the security of the current repository.” |
| Preparing a commit | Content staged with git add, with the previous version as context. | “Review the staged changes before I commit.” |
| Reviewing a branch or PR | A local comparison of versions, including deleted files and removed safeguards. | “Review this branch's changes against main.” |
| Checking someone else's skill or code | A specified directory, instruction file, supported archive, or public GitHub repository. | “Review this package, including prompt injection and obfuscation.” |
For branch comparisons, specify a base that exists in your repository. A PR URL alone does not replace the local data required for comparison. Scans run on demand; the skill does not install Git hooks or a CI pipeline.
Security checks
| Area | What the agent checks |
|---|---|
| Access and user data | Authentication, authorization, access to individual records, and tenant isolation. |
| Injection and code execution | SQL/NoSQL, shell commands, deserialization, templates, and unsafe use of model output. |
| Network, browser, and files | SSRF, XSS, session handling, uploads, file paths, and symbolic links. |
| Secrets and data flow | Hardcoded keys, sensitive data in logs, suspicious data transfers, cryptography, and TLS in context. |
| Dependencies and configuration | Package sources, installation scripts, CI configuration, broad permissions, and execution of code from untrusted sources. |
| AI and agents | Prompt injection, system prompt extraction, memory poisoning, tool misuse, and overly broad skill triggers. |
| Obfuscation and package contents | Unicode Tags, invisible characters, homoglyphs, encoded content, hidden scripts, actual file types, and document contents. |
| Logic and resources | Expensive or unbounded operations, relevant race conditions, and logic flaws in sensitive workflows. |
A pattern match is a starting point for analysis. The agent must check the data source, sensitive operation, existing safeguards, and potential impact. Using an HTTP function, subprocess, or JavaScript does not by itself establish a vulnerability.
Full scope: 18 security check areas · 8 application checks.
How reports are produced
| Stage | Result |
|---|---|
| Preparation | A copy of the code to review and a manifest listing files, versions, changes, and omissions. |
| Analysis | Pattern searches, contextual reading, and comparison of both sides of changes. |
| Assessment | Findings with locations, evidence, severity, confidence, impact, and suggested fixes. |
| Validation | Checks of report structure, its connection to the manifests, and preservation of information about omissions. |
The report shows what was reviewed and what could not be reviewed. An incomplete review remains marked as partial. The validator checks data consistency; the agent evaluates the accuracy of findings.
The default format is Markdown in the user's language. JSON is available on request. An optional 0–100 score is available only for a complete audit of an individual skill; it does not represent an application's percentage of security.
Verified by tests
105 out of 105 tests passed, with no skipped tests, in the run on November 15, 2026, using macOS 26.5.2 and Python 3.9.6. The tests included valid inputs as well as deliberately corrupted or manipulated samples.
The tests cover the shared 2.3.0 package. Codex CLI 0.166.0 recognized the portable manifest and the plugin in the marketplace catalog. Claude Code 2.1.325 validated the manifests and installed the plugin in a temporary profile, discovering exactly one skill. Standalone installation was tested for both agents, including updates and write failures.
| Test group | Count | What was verified |
|---|---|---|
| Code preparation and Git | 31 | Staged content differing from working files, removed safeguards, a repository with 2,400 files, limits, and no execution of test commands embedded in Git configuration. |
| Reports and manifests | 14 | Rejection of substituted manifests, fabricated locations, hidden omissions, and contradictory claims of completeness. |
| GitHub and search | 4 | Archive comparison against a commit tree using simulated data, URLs, hidden files, and data containing NUL bytes. |
| Archives | 34 | CRC, duplicates, unsafe paths, additional data, decompression limits, cleanup after errors, and compatibility with valid archives. |
| Package and regressions | 7 | Links, path portability, completeness of copied installations, and correct reporting of test failures. |
| Distribution and installer | 15 | Shared source, manifests, client discovery of the plugin, installation, updates, recovery after SIGKILL, locking against concurrent changes, and rejection of invalid recovery state. |
View the full 105-test run log. The shared skill also passed OpenAI's official structure validator, and the root manifest passed validation against the official Agent Plugins 1.0.0 schema.
The regression suite also includes 51 cases from the threat catalog: 48 fully checked automatically, 1 partially checked, and 2 still requiring agent assessment, plus 23/23 additional cases. These are part of the test suite above, not additional tests to add to its total.
A full audit performed by the model in a fresh session of each agent, downloading and updating the published package through the marketplace, and operation on Linux and WSL still require separate verification. The 105/105 result covers the tested scripts, rules, installer, and the client checks described above; it does not establish a complete end-to-end model run.
Requirements and limitations
- Agent: Codex or Claude Code. Claude support is intended for Claude Code.
- Tools: Python 3.9+ without additional libraries, Git 2.36+ for Git scopes, and ripgrep or a compatible search tool provided by the host. Additional file type and document tools are described in the documentation.
- System: Verified on macOS. Linux and WSL require separate testing; native Windows is not supported.
- Type of analysis: Static review. Git history, runtime application tests, container image scanning, and current dependency vulnerability databases are not checked automatically.
- Privacy: Reports must mask secrets as
[REDACTED]. Code that is read may be sent to a cloud model according to the agent's settings; local files do not imply offline analysis. - Untrusted code: Review it in a controlled session without loading its instructions or configuration as the auditor's settings. The skill does not undo configuration already loaded by the host.
The prescribed procedure does not execute the code under review. Incomplete reads, missing tools, or exceeded limits must be disclosed in the report. Results support security review and do not replace application testing or deployment decisions.
Documentation
| Resource | Contents |
|---|---|
| Shared instructions | Audit procedure for both agents. |
| Tests and test requirements | Regressions, installation, updates, and client integration. |
| Review scopes | Repository, staged changes, branch, archive, limits, and omissions. |
| Check catalog | Skill security, AI, and obfuscation rules. |
| Application checks | Review of access, data, and sensitive operations in code. |
| Report format | Findings, coverage, manifests, and validation of report contract 1.2. |
| Public GitHub repositories | Downloading a specified version and verifying its scope. |
SKILL.md, the reference files, and the scripts have a single source. Host permission differences are described in the shared instructions; OpenAI interface metadata is in agents/openai.yaml.
Author and license
Author: rkolcz. Skill version: 2.3.0.
Copyright 2026 rkolcz. All rights reserved. This product is covered by the custom Security Scanning commercial license. The root LICENSE covers the package for both agents and includes the author's details.
The license covers one purchaser or organization and allows reviews of both your own projects and client projects. Reports may be delivered to clients and used commercially. Sharing, publishing, or reselling the skill itself, including modified versions, requires separate permission from the author, subject to the exceptions and rights described in the license. See LICENSE for the complete terms.