Open Source
Open Source

Codex Security: CLI and TypeScript SDK, Apache-2.0 Code

An open-source resource note built around a single repository, openai/codex-security. It reads the GitHub API fields, the repository README, and the LICENSE text word for word, separates the Apache code license from Trusted Access for Cyber, and treats the star count as the cumulative total returned by the API. No benchmark numbers or audit workflow are added.

Published November 14, 20269 min read
<!-- codex-security-opensource-resource | open-source | Codex Security: CLI and TypeScript SDK, Apache-2.0 Code -->

This is an open-source resource note built around a single code repository. It reads three things in openai/codex-security, a repository under the official openai organization: the fields returned by the GitHub API, the README inside the repository, and the LICENSE file at the repository root. What each can prove, and what it cannot, is the only question here. It is not a comparison between competing security tools, not a step-by-step audit workflow, and not a place for conclusions the source files never wrote. Every field and quoted line comes from a first-hand reading of the interface and files as of November 13, 2026. The repository homepage is https://github.com/openai/codex-security and can be fetched directly. Fixing the scope at one repository and three sources matters: the moment leaderboard ranks, growth rates, or download counts are mixed in, numbers that are not present inside the repository, the verification chain breaks. The point of a source note is to draw a checkable boundary, not to make the volume of attention sound lively, and that boundary is what the rest of this article keeps returning to.

Repository identity: reading the fields one by one

The GitHub API returns HTTP 200 for the repos endpoint of openai/codex-security. The full_name field reads, verbatim, openai/codex-security, under the official openai organization. fork is false, archived is false, private is false, and default_branch is main. These state fields pin down the object: an independent primary repository owned by the organization, not a derived copy, not archived, not private, with a clearly named default branch. Every later citation lands on this one definite object, so it cannot be confused with a similarly named product from the same organization. Determinacy at the field level is the ground the rest of the article stands on.

The measured stargazers_count is 12,427, always written here as the API-measured 12,427 stars. The scope needs a note: 12,427 is the current cumulative total returned by the interface, accumulated since creation, not the number added on any single day. Reading it as more than ten thousand stars in one day misreads the field, and created_at rules that out: 2026-07-13T22:00:13Z means the repository was created on July 13, 2026, not thrown together on the day of reading. A repository that has existed for months cannot have its cumulative star count recast as a one-day increase. The measured forks_count is 978, also cumulative. language is TypeScript, fitting a repository whose main body is a TypeScript SDK.

Two more time fields matter. created_at is 2026-07-13T22:00:13Z and pushed_at is 2026-11-12T21:10:11Z, so the most recent push happened on November 12, 2026, matching the November 12, 2026 date on the official page. As of November 13, 2026, the repository was still maintained rather than idle. license.spdx_id is Apache-2.0 and homepage is developers.openai.com/codex/security. The description reads, verbatim, OpenAI's Codex Security CLI and TypeScript SDK for finding, validating, and fixing security vulnerabilities. npm: @openai/codex-security. One line names the command-line tool, the TypeScript SDK, the three actions of finding, validating, and fixing, and the npm package name.

The official page https://github.com/openai/codex-security was confirmed reachable rather than a 404. These fields come from that same first-hand read, not a third-party leaderboard or press account; where a field gives no quantity, this article invents none. The identity layer answers who owns it, when it was created and last touched, its main language, its license identifier, and its cumulative stars and forks, each traceable to a concrete field. Rankings, growth, and attention are not part of that answer.

Toolchain shape: two entry points, one loop

The README opens, verbatim, Codex Security helps you find, investigate, and fix vulnerabilities in your code., while the description speaks of finding, validating, and fixing security vulnerabilities. Together they close a loop: find a vulnerability, validate it, then fix it. This article reads that loop strictly as self-description, and infers no count of vulnerabilities found, no fix rate, and no effectiveness figure, because none appears inside the repository. Separating a self-description that can be read from outcomes that cannot be checked is the basic posture here.

The next sentence reads, verbatim, Use its security workflows in Codex, or automate them with the @openai/codex-security CLI and TypeScript SDK. Two entry points: use security workflows inside Codex, or automate them with the CLI and TypeScript SDK from the package @openai/codex-security. This is why the CLI and TypeScript SDK sit side by side in the title: one npm package offers both a terminal interface and a library TypeScript code can import, sharing one package name and one set of capabilities. A reader need not choose between a tool and a library, because the repository provides both forms.

On the command line, install is verbatim npm install --global @openai/codex-security, followed by two starting commands cs login and cs scan .: the first logs in, the second scans the current directory. The README also names outputs the command line can produce: drafting a SECURITY.md, exporting a threat model, emitting results in SARIF, JSON, and CSV, integrating with GitHub Actions, and batch scans with Docker. These are named capability items, recorded here without inferring how broad or effective each is, and without arranging them into a workflow a reader is expected to follow.

On the library side, the import reads, verbatim, import { CodexSecurity } from "@openai/codex-security";. The runtime requirement reads, verbatim, Requires Node.js 22.13.0+ within 22.x, or Node.js 24.x or 26.x, and Python 3.10+. Node.js within 22.x needs to be 22.13.0 or newer, or one may use 24.x or 26.x, with Python 3.10 or newer. The figures Node.js 22.13.0 and Python 3.10 are kept verbatim rather than softened into a vague recent version or extended with invented bounds. Version requirements are the first thing to check before starting, and quoting the original is safer than paraphrasing it.

The provider sentence reads, verbatim, Scans support OpenAI, Amazon Bedrock, OpenRouter, Fireworks AI, and Cheaper Inference., naming five. A boundary sentence reads, verbatim, Some cybersecurity requests and protected findings require Trusted Access for Cyber, noted as related to Daybreak opt-in. This is a condition at the capability-access level, not the license on the code. Recording it heads off a common misreading: code being obtainable does not mean every capability needs no application. The original names a condition, so this article keeps it as written, without explaining how to apply for it or softening it.

License: the LICENSE file is the only authority

The root LICENSE is the full standard Apache License, Version 2.0 text, 10,926 characters long. On that basis this article makes one confident statement: the code is licensed under the Apache License. What is confirmed is the code layer using Apache 2.0, because the LICENSE text says so directly. Any claim beyond that layer, such as a broad promise that the entire repository and everything unwritten about it is free of all restrictions, is outside what the text covers and is not made here. Licensing is where drifting beyond the source on the strength of an impression does the most damage, so the wording is narrowed to the one sentence the file supports.

One detail is easy to miss: the README has no section headed License anywhere in its full text. It covers installation, commands, runtime requirements, providers, and restricted capabilities, but has no passage declaring a license. The license therefore cannot be inferred from the README; the root LICENSE is the only authority. The API field license.spdx_id returns Apache-2.0 and corroborates the file, but the file itself governs, not the field label standing in for it. A field may corroborate while the file governs, which is why the LICENSE text is stressed instead of a loose statement that the repository is Apache 2.0.

Keeping code licensing separate from Trusted Access for Cyber matters just as much. The first says under what license the code is released; the second says some cybersecurity requests and protected findings need an additional access condition. One governs how code is licensed, the other how capabilities are invoked, on different layers. The README names the latter explicitly, so capabilities are not all open unconditionally. This article places the two side by side without explaining the condition or misreading it as a restriction on the code license.

What can and cannot be verified

Taken together, what can be quoted verbatim is clear. From the GitHub API: the owning organization, cumulative stars and forks, primary language, license identifier, creation and last-push timestamps, default branch, the not-forked, not-archived, not-private status, the description, and the homepage. From the README: the finding-validating-fixing loop, the two entry points of the CLI and TypeScript SDK, the npm package name, the login and scan commands, the SDK import, the Node.js and Python thresholds, the five providers, the output formats and integrations, and the Trusted Access for Cyber boundary. From the LICENSE: the full Apache License text. These three groups are the entire first-hand material cited here.

Equally important is what is absent: no performance benchmarks, no download or npm install counts, no statistics on vulnerabilities found or fix rates, and no field from which a star growth rate could be derived. The 12,427 figure stays a cumulative total and cannot be broken into an increase over an interval. This article neither fills those gaps nor turns the absence of a figure into a claim that the owner denied anything; the former is a boundary of the material, the latter would be speaking on the repository's behalf. A first-hand read lays out what can be checked and leaves the rest blank, with the blank itself part of the facts. Every claim above can be traced back to one of those three sources, and nothing outside them is asserted. Holding that line keeps the article from being pulled off course by lively numbers that the repository never supplied.

It helps to state how this article divides from others on the site. Tool comparisons, coding-agent security coverage, and the codebase audit workflow are handled separately in the code security audit tools comparison, the coding-agent security hotspot, and the codebase security audit SOP. Those deal with comparing tools or running an audit workflow; this article reads only the fields and the README and LICENSE text of the single repository openai/codex-security. It is also distinct from the openai codex harness article: same organization, different product, and no information from the other product is folded in.

FAQ

Q1: What is the openai/codex-security repository for? A1: A toolchain closing a loop of finding, validating, and fixing vulnerabilities. The README gives two entry points verbatim: use security workflows inside Codex, or automate them with the CLI and TypeScript SDK from the npm package @openai/codex-security. The command line has install, login, and scan commands, and the SDK imports CodexSecurity. What is verified here are those fields and lines, not how many vulnerabilities the tool has actually found.

Q2: What license does the repository code use? A2: The root LICENSE is the full standard Apache License, Version 2.0 text, so this article writes that the code is licensed under the Apache License. The README has no License section, so the LICENSE file is the only authority; the API field license.spdx_id returns Apache-2.0 as corroboration. Claims beyond the code layer are not supported by the text and are not made here.

Q3: What are the runtime requirements? A3: The README requires, verbatim, Node.js 22.13.0+ within 22.x, or Node.js 24.x or 26.x, and Python 3.10+. Install with npm install --global @openai/codex-security, then cs login and cs scan . to start. The version figures are kept exactly as written rather than softened into a general statement.

Q4: How should the API-measured 12,427 stars be understood? A4: 12,427 is the current cumulative total returned by the interface, not a one-day increase. created_at is 2026-07-13T22:00:13Z, so the repository was created on July 13, 2026, not set up on the day of reading, and its cumulative star count cannot be recast as a single-day increase; forks_count is 978, also cumulative. As of November 13, 2026, the most recent push was on November 12, 2026.

Q5: Can every capability be used directly? A5: That cannot be inferred. The README states verbatim that some cybersecurity requests and protected findings require Trusted Access for Cyber, a capability-access condition distinct from the code being licensed under the Apache License. Code being obtainable does not mean every capability needs no application; the two statements are recorded side by side without explaining the condition on the repository's behalf.

Related reading

To move beyond a single repository, see the sandbox isolation comparison, which compares isolation approaches rather than reading one repository's fields and source text. Back within this article, what stays certain is those three places: the GitHub API fields, the README, and the LICENSE text. Anything those three do not provide is left blank rather than filled in.

This article is AI-assisted and human-edited. Last updated: 2026-11-14

FAQ

What is the openai/codex-security repository for?
A toolchain closing a loop of finding, validating, and fixing vulnerabilities. The README gives two entry points verbatim: use security workflows inside Codex, or automate them with the CLI and TypeScript SDK from the npm package @openai/codex-security. The command line has install, login, and scan commands, and the SDK imports CodexSecurity. What is verified here are those fields and lines, not how many vulnerabilities the tool has actually found.
What license does the repository code use?
The root LICENSE is the full standard Apache License, Version 2.0 text, so this article writes that the code is licensed under the Apache License. The README has no License section, so the LICENSE file is the only authority; the API field license.spdx_id returns Apache-2.0 as corroboration. Claims beyond the code layer are not supported by the text and are not made here.
What are the runtime requirements?
The README requires, verbatim, Node.js 22.13.0+ within 22.x, or Node.js 24.x or 26.x, and Python 3.10+. Install with npm install --global @openai/codex-security, then cs login and cs scan . to start. The version figures are kept exactly as written rather than softened into a general statement.
How should the API-measured 12,427 stars be understood?
12,427 is the current cumulative total returned by the interface, not a one-day increase. created_at is 2026-07-13T22:00:13Z, so the repository was created on July 13, 2026, not set up on the day of reading, and its cumulative star count cannot be recast as a single-day increase; forks_count is 978, also cumulative. As of November 13, 2026, the most recent push was on November 12, 2026.
Can every capability be used directly?
That cannot be inferred. The README states verbatim that some cybersecurity requests and protected findings require Trusted Access for Cyber, a capability-access condition distinct from the code being licensed under the Apache License. Code being obtainable does not mean every capability needs no application; the two statements are recorded side by side without explaining the condition on the repository's behalf.

Related

Open Source

LingBot-Touch: Code Apache-2.0, Weights Elsewhere

Robbyant/lingbot-touch reached 238 stars within days of creation, with Apache-2.0 in both the license field and the badge alt text, yet the model weights are not in the source repository: the README states verbatim that model files are hosted separately and access requires an account with permission to the model repository. This piece reads only API fields, README wording, and the LICENSE text: created November 9, 2026, last push November 11, main language Python, default branch main, not a fork and not archived. It also covers how three heterogeneous tactile sensors feed one representation and how the two model variants differ, without importing scores, download counts, or star totals the repository never states. As of November 12, 2026, code licensing and gated weights remain two separate questions.

Nov 13, 20269 min read
Open Source

Tencent Open-Sourced Simultaneous Interpretation: 3s Lag

Tencent Hunyuan open-sourced the simultaneous interpretation model Hy-SimulMT on October 26 (30B-A3B, about 30.39B parameters per the HF weights with 128 experts and 8 activated per token): real-time Chinese-English speech translation with StreamLAAL of about 3 seconds on the RealSI long-form benchmark (the project notes this is not the wait before the first translated word), a GEMBA of roughly 85-86 per the official comparison chart (scored by GPT6.1-Sol) beating the compared open models and 4 of 5 commercial APIs, and lower latency than six CATTI-certified interpreters (4.5/4.7 seconds) across an 18-recording small-scale test. Key correction: the base checkpoint outputs translated text only — speech output and voice cloning need a separate TTS service (the official recipe uses Qwen3-TTS), and the 3-second figure excludes synthesis. License in one line: Apache-2.0 per the LICENSE file with third-party exceptions (only Qwen3-ASR-0.6B) and no extra restrictions, while GitHub shows NOASSERTION due to the custom license filename. Six builds (BF16/FP8/GGUF/Interleave), an official anchor of about 57 GiB on a single H200 for BF16, GGUF minimums unpublished; tooling includes SimulInfer/SimulTrain, llama.cpp support is PR #31198 (unmerged), and single files cap at 30 seconds.

Oct 28, 202612 min read
Open Source

Reflection AI Beam 501B: Can Four B200s Carry 1TB of Weights?

Reflection AI's first open-weight model Beam actually landed on 2026-10-16: repository huggingface.co/reflection/Beam-1.0-501B-A23B (plus FP8 and NVFP4 variants) under Apache 2.0, with the model card marked informational only and non-binding. The timeline matters: the 10-05 announcement came with only a waitlist and a promise of weights this month, a benchmark table update on 10-08, no public repository on the HF org page as of 10-13, then weights, Hugging Face's own blog post and an AWS SageMaker path all landing on 10-16 - eleven days from promise to delivery. Architecture: a sparse MoE with 501B total parameters, 23B active per token (about 4.6 percent), 52 layers, 256 experts with 8 routed per token, text in and text out; the 1M effective context stretched during midtraining and the 256K RL rollout length are two different figures that must not be merged, and the serving-side context ceiling and max output are unpublished. Reasoning depth has five levels (low to max) and the weights bundle a DFlash2 speculative-decoding drafter. The hardware bill is the core: BF16 at about 1TB, NVFP4 at about 280GB with Hugging Face saying two B200s fit barely and recommending four B200s (720GB); the smallest GGUF is about 103GB and is a RAM requirement first. All benchmarks are Reflection self-reported with zero independent replications: SWE-bench Verified 81.6 percent (FP8 80.2, NVFP4 80.7), Terminal-Bench v2.1 80.1, GPQA Diamond 90.6, AIME 2026 97.7, HLE without tools 36.0; Artificial Analysis has indexed it at Intelligence Index 30, second among US open-weight models, relayed secondhand. Ollama offers day-zero free access but ollama run beam pulls beam:cloud, its hosted inference - local runs need an explicit local GGUF. Mirror CLI (Rust) ships as an open agent harness alongside 40-plus partner integrations; the self-distillation recipes remain unpublished, and the FT report that Nvidia is in talks to acquire or deepen its stake in Reflection is unconfirmed by both parties.

Oct 18, 202610 min read