zizmor: workflow security is checked before the code ever runs
zizmor is a static security analyzer for CI/CD files — GitHub Actions, Dependabot, pre-commit. It reads the YAML before anything executes, finds attack paths, and runs offline by default, with no account and no network. It is a ready-made binary with zero Python dependencies, zero tokens in the model and zero money. In our test, an 11-line workflow with deliberate mistakes produced five findings and exit code 14 in 0.18 seconds.

What it does
Every time you push code to GitHub, some YAML files decide what runs, with which permissions and with which secrets. Those are GitHub Actions workflows. They are the automatic machinery that builds, tests and ships your application. The problem is that the same files are also an attack surface. One mistake on one line can give outside code access to your keys, or run a command written by someone outside your team.
zizmor reads those files and looks for such mistakes. It does not run the workflow, sends nothing, asks for no account. It is static analysis: it reads the text, understands its structure, and compares it against known patterns of bad practice. It shows the problem before it can do damage.
The categories it catches have names. Template injection: when a message or title from outside goes straight into a shell command, an attacker can write their own command. Unpinned uses: when you call an action by a tag such as v4 instead of a full commit hash, and the tag changes hands. Dangerous triggers: when a workflow listens to an event that brings code from an external fork, such as pull_request_target. Excessive permissions: when the workflow asks for more rights than it needs. And artipacked: when checkout leaves credentials inside the copy and they end up in an artifact.
These are not theoretical. They are how real breaches have happened in public repositories. The tool does not ask you to learn the theory on day one. It gives you the finding, the rule, and often a suggested fix.
One point to clear up: zizmor is not an MCP server. It does not speak to the model, loads no tool schema, takes no room in any session. It is a command-line tool. You run it from the shell, it reads files, prints results, exits. That distinction changes how you wire it into your flow.
The zizmorcore team maintains it. The repository has existed since 19 August 2024 and today holds 6,804 stars, 274 forks and an MIT licence. Its last activity is today, 28 October 2026. It is sponsored by Grafana Labs, Trail of Bits and Kusari (repository).
On PyPI the latest stable version is 1.31.1, published on 15 October 2026, using the wheel’s upload_time field — not another field, since metadata often shows different dates. There is also a 1.32.0-rc1 from 23 October 2026 (PyPI).
Who it is for
For anyone who has a workflow on GitHub and has never read it line by line. That includes almost everyone using Actions, from teams with ten repositories to someone who set up their first CI two weeks ago.
The beginner gains the simplest thing: one command, a list of mistakes in plain language, and often an automatic fix. They do not need to know what pull_request_target is to understand that their workflow listens to external pull requests and prints their title inside a command. The tool tells them in words.
The professional gains something sharper: a gate that fails the build when someone adds a dangerous line. It does not replace judgement; it stops judgement getting lost inside a fifty-line diff. It shines when workflows are many or when many hands touch them, because there the eye gets tired and a v4 tag slips through.
It is not for you if you do not use GitHub Actions at all. It is also not for you if you want a tool that proves a workflow is safe. zizmor shows mistakes in known patterns. It does not issue a security certificate.
Install
The installer is a Python package, but the program you execute is a pre-compiled binary. Python is needed only to download and place the file. After installation, zizmor runs without Python. This matters: you are not adding another language to your machine for a job that takes ten seconds.
The installer requires Python 3.11 or newer. The package’s requires_dist field on PyPI is empty — zero Python dependencies. It pulls no libraries, builds no package tree, leaves nothing behind except the binary itself.
The tested route, from any folder:
uv tool install zizmoruv installs the tool into an isolated environment and adds the zizmor command to PATH. No virtual environment, no activation. If you already have uv, it is one command and you are done.
The vendor documents other routes: Homebrew, cargo install from source, and a direct binary from the releases. That is a vendor claim. We only ran the uv route, and I will not give you instructions for steps I did not execute.
Using it per LLM
The same rule applies here as for any command-line tool. You do not declare a server. You do not write an mcpServers entry. You do not restart so the assistant can see it. zizmor is not an MCP server, and if you put a server schema on it you will be writing a lie.
What you do is call it from the shell, either before you touch a workflow or inside CI as a gate. The same command works everywhere, because it does not depend on the assistant. That is the advantage: you do not lose the investment when you change tools.
From the root of your repository:
zizmor .github/workflows/That command reads every file in the folder and prints the findings. If you want it to fix whatever it can fix on its own, on one specific file:
zizmor --fix .github/workflows/ci.ymlBefore you let an assistant write or change a workflow, the gate is simple: run zizmor first, see what it says, and only then proceed. The same applies in reverse — after the assistant finishes, run the check again. You do not need to explain to it what zizmor is. You give it the command and the exit code.
Claude Code
You write the convention into CLAUDE.md: before you change a workflow, run zizmor. And if you want to do it by hand, from the root:
zizmor .github/workflows/Codex CLI
The same shape, in an AGENTS.md file. No settings section, just the convention and the command:
zizmor .github/workflows/Cursor
Here too, one instruction in the project rules, no server declaration. For a manual check:
zizmor .github/workflows/In CI
This is the strongest use, because the exit code becomes a gate. You add a step before the build:
zizmor .github/workflows/If the step exits with 14, the pipeline stops. If it exits with 0, it proceeds. That way nobody can pass a dangerous change without the tool seeing it. No server declaration is needed here either; it is a plain shell command.
Where it wins, where it loses
It wins first on needing nothing. No account, key, connection or config file. You run it and it answers. That is rare among security tools, which usually want platform integration or a service connection.
It wins on running offline by default. It prints this itself on execution: “zizmor is running in offline mode by default; some audits and auto-fixes will not be available”. So it does not send your files anywhere. That matters on restricted networks, and for anyone who does not want the names of internal flows on a foreign server.
It wins on speed. In our test, analysing an 11-line workflow took 0.18 seconds. It can go into every commit without anyone noticing.
It wins on the exit code. 0 means clean, 13 means warnings only, 14 means errors. That three-way split is useful: you decide whether to break the build on warnings or only on errors. The gate is yours, not the tool’s.
It wins on fixing, too. Two of the five findings in the test were fixed automatically with –fix. It does not solve everything, but it saves the mechanical work.
Now where it loses.
First, it is a static analyzer. It reads text. It does not execute the workflow, does not see what it does in practice, does not know whether an action behaves as promised. It can find a known pattern, but it cannot guarantee there is no other, unknown one. Coverage is only as good as the rules it knows.
Second, coverage may change in online mode. In online mode some audits pull action metadata from GitHub. The vendor says it is optional. We worked offline. We did not test online, so we do not know how much more it sees or how much less. That is a caveat, not a flaw: the tool is useful offline too.
Third, it wants YAML familiarity to fix what is not fixed automatically. The finding comes with an explanation, but the change to the file is still your work.
And the honest caveat we should not hide: a security tool does not make a workflow safe by itself. It shows mistakes, it does not erase them. If someone wants to insert a dangerous line, they can find a way the rule does not catch. zizmor reduces the surface; it does not eliminate it.
Testing
I wrote an 11-line workflow deliberately full of mistakes. It had pull_request_target as the trigger, a run step printing github.event.pull_request.title, and an actions/checkout@v4 without pinning to a hash. That is three of the most classic mistakes at once.
I ran the command from the repository root:
zizmor .github/workflows/It exited with code 14 in 0.18 seconds. Five findings, with the right colour in the terminal: template-injection at high confidence with an automatic fix, unpinned-uses at high, dangerous-triggers at moderate, excessive-permissions at moderate, and artipacked at low with an automatic fix.
Then I tried –fix on the same file:
zizmor --fix .github/workflows/ci.ymlIt fixed the two that had automatic fixes on its own. The other three were left for a human hand. That is correct: you do not expect a tool to know whether you want to change a trigger or permissions.
And then the clean one. I made a workflow without any of those mistakes and ran it. It exited with code 0 and printed “No findings to report. Good job!”. So the gate works in both directions: it stops when there is a problem, it passes when there is not.
What I did not measure, and I say so. I did not measure memory: a number from estimation would be invention. I did not test online mode. Any number not stated here is not ours.
The repository figures — 6,804 stars, 274 forks, MIT licence, last push 28 October 2026, first created 19 August 2024 — and the PyPI figures — 1.31.1 on 15 October 2026, 1.32.0-rc1 on 23 October 2026 — come from our research, as GitHub and PyPI report them.
What it costs
The model. Our own figure: zero tokens. zizmor is not an MCP server, it loads no tool schema, it takes no room in any session. You pay nothing in the assistant’s context, whether you use it or not.
The local footprint. Our own figure: the Windows wheel is 9.03 MB, or 9,029,164 bytes. The installed executable is 24.9 MB on disk, or 24,859,648 bytes. Zero Python dependencies. No process left running after execution. I did not measure memory, for the reason I gave.
In money, nothing. MIT licence, no key, no service, no account, offline by default. There is no paid tier and no telemetry we know of.
Vendor claim, not our own figure: that online mode is optional and that some audits pull action metadata from GitHub. We read that in the documentation and saw it in the offline message. We did not test it.
Verdict
Score: 8.7/10.
zizmor solves a problem most people do not know they have: their workflows are code with permissions, and nobody reads them. It solves it without an account, without a network, without dependencies, with one command. In our test: five findings in 0.18 seconds, exit code 14, two automatic fixes, and exit 0 on a clean file.
It keeps or loses the remaining points for three reasons. It is a static analyzer, so it sees patterns and not behaviour. Coverage may change in online mode, which we did not test. And it does not make a workflow safe by itself — it shows mistakes, it does not erase them.
Where we recommend it without hesitation: in every repository with GitHub Actions, as a step before the build and as a habit before you touch a workflow. There it is the cheapest security gate you can put in place today.
Sources: repository https://github.com/zizmorcore/zizmor · audits documentation https://docs.zizmor.sh/audits/ · PyPI package https://pypi.org/project/zizmor/