OpenAI Codex Hits 1.6M Weekly Users — Launches Dedicated Security Vulnerability Agent

OpenAI published Codex adoption numbers following the GPT-5.3-Codex release: 1.6 million weekly active users, with enterprise deployments at Cisco, NVIDIA, Ramp, Rakuten, and Harvey. Alongside the usage data, they launched Codex Security — a dedicated application-security agent separate from Codex’s general coding functionality.

Why a Separate Security Agent

The decision to ship a standalone security agent rather than fold it into general code review is deliberate. Security analysis requires different validation logic:

• A code quality finding that’s wrong is annoying
• A security finding that’s wrong (false positive) creates alert fatigue and trains engineers to ignore warnings
• A security finding that’s missed (false negative) is a production vulnerability

Codex Security has a validation step where the agent confirms a vulnerability is real before surfacing it. This step is specifically designed to reduce the false-positive rate that has historically undermined SAST (static application security testing) adoption. Teams stop reading SAST output when it cries wolf too often.

The agent covers:

• Injection vulnerabilities — SQL injection, command injection, template injection
• Authentication issues — broken auth flows, session management problems
• Sensitive data exposure — hardcoded secrets, insecure storage patterns
• Dependency vulnerabilities — CVEs in your package lock files
• API security — rate limiting gaps, missing authorization checks

The 1.6M Number in Context

1.6M weekly active users is a notable scale milestone, but the enterprise customer list is the more structurally interesting data point. NVIDIA and Cisco are organizations with large, complex codebases and strict security requirements. Their production deployments signal that enterprise-grade reliability is achievable with current Codex quality — a threshold that matters for procurement decisions at other large organizations.

The Ramp and Harvey deployments are worth noting separately. Financial infrastructure (Ramp) and legal AI (Harvey) both operate under regulatory requirements that make AI-generated code a higher-stakes proposition. These aren’t experimental pilots.

Codex vs. Claude Code Review vs. Cursor

The competitive picture for AI-assisted development is consolidating around three serious options:

• Codex — strongest API ecosystem integration, 1.6M WAU scale, new security specialization
• Claude Code Review — parallel agent architecture, REVIEW.md customization, $15–$25/review pricing clarity
• Cursor Automations — event-driven always-on model, MCP integration, PagerDuty incident response

These aren’t mutually exclusive. Teams are increasingly using Codex for code generation, a review tool for PR analysis, and an automation layer for ongoing monitoring. The question isn’t which one to pick — it’s which combination fits your workflow and budget.

Access

Codex Security is available now in the OpenAI API and in ChatGPT for Enterprise users. The security agent runs as a separate workflow from standard Codex completions — you opt in per-repository.

Source: OpenAI — Codex usage data, March 2026