An application security agent developed by OpenAI. It analyzes an application’s codebase and handles vulnerability discovery, verification, and remediation end to end. It surfaces weaknesses while explaining how the implementation works in natural language, verifies whether each finding is actually exploitable through sandboxed execution to eliminate false positives, and presents the results ranked by severity. Its predecessor was “Aardvark,” a research agent announced in October 2025; the product officially launched in March 2026, and an open-source (Apache 2.0) CLI and TypeScript SDK were released on July 28, 2026. It runs on Mac, Windows, and Linux.
Key Features
- End-to-end discovery, verification, and remediation: Beyond scanning code for vulnerabilities, it can automate patch generation and application, all the way to creating GitHub pull requests
- Sandbox verification to eliminate false positives: Detected vulnerabilities are actually verified in an isolated environment before being reported, so you are far less likely to be buried under the mass of false positives typical of static analysis tools
- Natural-language explanations: It points out weaknesses while explaining how the implementation behaves in plain language, making the reasoning behind each finding easy to understand and useful for reviews and learning
- Scan history comparison: It compares results against past scans to automatically classify findings as new, recurring, or resolved, making continuous security progress easy to track
- Bulk scanning and CI/CD integration: It supports bulk scanning of multiple repositories via Docker containers, and non-interactive execution with an API key lets it plug into CI/CD pipelines
- Deep mode for exhaustive inspection: A deep inspection mode runs for up to 96 hours with configurable worker counts and stop conditions to examine a codebase thoroughly
Pricing
| Offering | Price | Notes |
|---|---|---|
| CLI / TypeScript SDK | Free (open source, Apache 2.0) | A ChatGPT account or an OpenAI API key is required to run inference |
| API key usage | Usage-based (depends on token consumption) | Token usage and estimated cost are shown per scan; a budget cap can be set with --max-cost |
| ChatGPT plan integration | Available to Pro / Business / Enterprise / Edu | See official information for details |
Pricing is as of August 2026. Scan costs vary with the complexity of the analysis. For the latest information, check the official repository and the documentation.
Pros & Cons
✅ Pros
- A single agent covers everything from discovery to patch and pull request creation, dramatically reducing remediation effort
- Findings are demonstrated in a sandbox before being reported, cutting the time spent triaging false positives
- The reasoning behind each finding is explained in natural language, making decisions easier even for teams without a dedicated security engineer
- The CLI / SDK are open source, giving you plenty of freedom for CI/CD integration and custom tooling
- The scan comparison feature lets you continuously track vulnerability remediation progress
⚠️ Cons
- Running inference requires a ChatGPT account or an API key, and costs scale with usage
- Scan cost depends on the agent’s reasoning volume rather than lines of code, making costs hard to estimate in advance
- Some features are said to require OpenAI approval (Trusted Access for Cyber)
- As a new tool launched in 2026, it has a shorter operational track record than established SAST products
- Exhaustive inspections such as deep mode can become large in both runtime and cost
Comparison with Similar Services
| Comparison | Codex Security | Snyk Code | GitHub Advanced Security | Semgrep |
|---|---|---|---|---|
| Provider | OpenAI | Snyk | GitHub | Semgrep |
| Detection approach | AI agent + sandbox verification | Static analysis + AI | CodeQL (static analysis) | Rule-based static analysis + AI |
| Auto-fix | Patch generation + PR creation | Fix suggestions | Copilot Autofix | Assistant suggestions |
| False-positive handling | Proven in a sandbox before reporting | Suppressed by proprietary engine | Depends on query precision | Suppressed via rule tuning |
| Open source | CLI / SDK under Apache 2.0 | Closed | CodeQL partially open | Engine is open |
| Pricing | Usage-based + ChatGPT plans | Free tier / paid plans | GitHub paid plans | Free tier / paid plans |
Who Is It For
- Startups and small teams without a dedicated security engineer that want to handle vulnerability discovery through remediation entirely within the development team
- Developers tired of triaging false positives from existing static analysis tools who only want to see verified findings
- DevOps engineers who want to build AI-powered security review into their CI/CD pipelines
- Organizations with many repositories that want a cross-cutting inventory via container-based bulk scanning
- Anyone who wants to learn secure coding while understanding the reasoning behind each finding through natural-language explanations
Summary
Codex Security is OpenAI’s security agent that automates vulnerability discovery, verification, remediation, and pull request creation end to end. Its two biggest differences from conventional static analysis tools are false-positive elimination through sandboxed execution and natural-language explanations. The CLI / SDK are published as open source, so a realistic adoption path is to scan one of your repositories first to gauge finding quality and cost, then expand to CI/CD integration and bulk scanning if it proves worthwhile.