Open source guide · Updated 2026-09-25

The Best GitHub Repos for AI Builders and Security Teams in 2026

Most best GitHub repos lists are written for people learning to code. This one is for the people who have to ship AI features and then defend them to a customer security questionnaire or an auditor. Every repo below is something we would put in front of a founder, a platform engineer or a GRC lead, and every entry says what it is for, how to try it in one command, which controls it helps you evidence, and where it falls short.

We checked every repo against its own README and the GitHub API on the date at the top of this page. Star counts are live figures from that check, not rounded marketing numbers. Three repos that would have made this list a year ago are gone, and the section at the end explains why, because an unmaintained dependency is a finding in its own right.

Stars and activity from the GitHub API, checked 2026-09-25

The list at a glance

#RepoWhat it is forStars
1Claude CodeAnthropic’s agentic coding tool for the terminal, IDE and CI.147,986
2OpenAI Codex CLIOpenAI’s open source coding agent that runs locally in your terminal.126,363
3MCP reference serversThe reference Model Context Protocol servers maintained by the MCP steering group.90,582
4n8nSelf-hostable workflow automation with native AI agent nodes and 400+ integrations.205,896
5OllamaThe simplest way to run open weight models on a laptop or server.181,654
6vLLMThe high-throughput inference engine behind most serious self-hosted LLM deployments.92,645
7LiteLLMAn AI gateway that puts 100+ model providers behind one OpenAI style API.59,579
8LangfuseOpen source tracing, evals and prompt management for LLM applications.35,026
9PromptfooEvals and automated red teaming for prompts, agents and RAG, run from the command line or CI.25,428
10garakNVIDIA’s LLM vulnerability scanner, a probe library for model level weaknesses.9,350
11NeMo GuardrailsProgrammable input, output, dialog and retrieval rails for LLM applications.7,188
12PresidioDetect, redact and anonymise personal data in text, images and structured data.11,028
13GitleaksFast secret scanning for git history, directories and stdin.29,472
14TrivyOne scanner for vulnerabilities, misconfigurations, secrets and SBOMs across images, code and clusters.38,068
15SemgrepFast static analysis with rules that look like the code they match.16,756
16zizmorStatic analysis for GitHub Actions workflows.6,574
17ProwlerOpen source cloud security posture checks mapped to SOC 2, ISO 27001, PCI DSS and more.14,863
18CheckovPolicy as code scanning for Terraform, CloudFormation, Kubernetes, Helm and more.9,028
19Open Policy AgentA general purpose policy engine for authorisation, admission control and CI gates.12,270
20WazuhOpen source SIEM and XDR with file integrity, configuration assessment and compliance mapping.16,992

Swipe sideways to compare columns

Coding agents and agent plumbing

The tools that now write and run a meaningful share of production code, and the protocol and workflow layers that connect them to everything else.

#1★ 147,986· TypeScript· last push 2026-09-25

Claude Code

anthropics/claude-code · Anthropic commercial terms (not an open source licence)

Anthropic’s agentic coding tool for the terminal, IDE and CI.

Claude Code reads your codebase, edits files, runs commands and handles git from natural language instructions. The repository is where Anthropic ships the plugins, examples and issue tracker, and it is one of the most starred developer tools on GitHub. It is on this list because it is where a lot of your code is now being written, which makes its configuration a security control whether you planned it or not.

Key features

  • Permission modes that decide which commands and edits need approval
  • Hooks that run your own checks before or after tool calls
  • MCP support for connecting internal tools and data
  • A plugin system with marketplaces, installable per project or per user

Try it

curl -fsSL https://claude.ai/install.sh | bash

Controls it helps evidence

ISO 27001 A.8.28 secure coding and A.8.32 change management, SOC 2 CC8.1, ISO 42001 Annex A AI system lifecycle

Honest take

The best general purpose coding agent we have used, and the hooks and permission settings are the part most teams skip. Commit a shared settings file with the permission rules you actually want, and treat plugins as third party code: Plugin4Shell in September showed that a pinned plugin is only as safe as the agent’s check that the pin landed.

Cons

Not open source, so you cannot audit or fork the CLI itself. Usage needs a paid Anthropic plan or API spend, and costs scale with how much autonomy you give it.

#2★ 126,363· Rust· last push 2026-09-25

OpenAI Codex CLI

openai/codex · Apache 2.0

OpenAI’s open source coding agent that runs locally in your terminal.

Codex CLI is a Rust agent that plans, edits and runs code on your machine, with the same agent available in the ChatGPT app, IDE extensions and the cloud. Unlike most commercial agents the CLI is genuinely open source under Apache 2.0, so your security team can read exactly what it does with your files and your shell.

Key features

  • OS level sandboxing of commands on macOS and Linux
  • Approval modes from read only to full auto
  • Works with a ChatGPT plan or an API key
  • MCP client support and a non-interactive mode for CI

Try it

npm install -g @openai/codex

Controls it helps evidence

ISO 27001 A.8.28 and A.8.32, SOC 2 CC8.1, ISO 42001 Annex A AI system lifecycle

Honest take

The sandbox is the reason to look at Codex if you care about blast radius: it constrains what a command can touch at the operating system level rather than trusting the model to ask first. Being Apache 2.0 also makes it the easiest coding agent to put through a proper supplier review.

Cons

Tuned for OpenAI models first. It shared the Plugin4Shell plugin pinning flaw with other agents, so check you are on a patched release.

#3★ 90,582· TypeScript· last push 2026-09-22

MCP reference servers

modelcontextprotocol/servers · Apache 2.0 (moving from MIT, older contributions remain MIT)

The reference Model Context Protocol servers maintained by the MCP steering group.

The Model Context Protocol is now the standard way to give agents tools and data, and this repo holds the small set of reference servers: filesystem, git, fetch, memory, time, sequential thinking and a test server called everything. It used to list hundreds of community servers. Those now live in the MCP Registry, and many older reference servers moved to an archive repo.

Key features

  • Reference implementations in TypeScript and Python
  • Examples of tools, resources and prompts
  • A test server that exercises every protocol feature
  • Links to the official MCP Registry for discovery

Try it

npx -y @modelcontextprotocol/server-memory

Controls it helps evidence

ISO 27001 A.5.15 access control and A.5.19 supplier relationships, SOC 2 CC6.1, ISO 42001 AI system inventory

Honest take

Read these to understand how MCP servers should be built, not to run them in production. The README says so directly: they are educational reference implementations, and you are expected to add your own safeguards for your own threat model. That sentence belongs in your AI acceptable use policy.

Cons

Deliberately small and not hardened. If you want a list of every MCP server out there, the Registry is now the place, and a listing there is not a security review.

#4★ 205,896· TypeScript· last push 2026-09-25

n8n

n8n-io/n8n · Sustainable Use License (fair-code, not OSI open source)

Self-hostable workflow automation with native AI agent nodes and 400+ integrations.

n8n is the most starred repo on this list, and for good reason: it lets operations teams wire LLM calls into real business processes with a visual editor, drop into code when they need to, and keep the whole thing on their own infrastructure. For regulated teams, self-hosting is the feature, because customer data never has to leave your environment to reach an automation vendor.

Key features

  • Visual workflow builder with JavaScript and Python code nodes
  • AI agent, memory and tool nodes built on LangChain
  • 400+ integrations plus generic HTTP and webhooks
  • Self-host with Docker or use n8n Cloud

Try it

docker run -it --rm --name n8n -p 5678:5678 -v n8n_data:/home/node/.n8n docker.n8n.io/n8nio/n8n

Controls it helps evidence

ISO 27001 A.5.23 cloud services and A.8.32 change management, SOC 2 CC6.1 and CC8.1, PCI DSS Requirement 6

Honest take

Excellent, and dangerous in exactly the way spreadsheets were: anyone can build a workflow that moves customer data to a model with a stored credential. Put workflows in source control, review them like code, and keep an inventory of which credentials each one uses.

Cons

The licence restricts offering n8n itself as a service, so check it before building a product on top. Some team features such as SSO and environments sit in paid editions.

Run, route and observe models

Self-hosted inference, a gateway to put in front of every provider, and the tracing you need before anyone asks what the model actually said.

#5★ 181,654· Go· last push 2026-09-25

Ollama

ollama/ollama · MIT

The simplest way to run open weight models on a laptop or server.

Ollama wraps model download, quantisation and serving into one binary with a local API, and it runs the major open weight families including Gemma, Qwen, DeepSeek and gpt-oss. For compliance teams it is the fastest way to keep a sensitive workload entirely on hardware you control, for example classifying documents that are not allowed to leave the building.

Key features

  • One command to pull and run a model
  • Local REST API plus OpenAI compatible endpoints
  • Runs on macOS, Windows and Linux, CPU or GPU
  • Modelfiles to pin a model, parameters and system prompt

Try it

curl -fsSL https://ollama.com/install.sh | sh

Controls it helps evidence

ISO 27001 A.5.34 privacy and protection of PII and A.8.20 network security, ISO 42001 data for AI systems

Honest take

The default is safe because it binds to localhost. The incidents come from people exposing port 11434 to a network so a colleague can use it, and the API has no authentication. If it needs to be shared, put it behind a gateway with auth, which is what the next two entries are for.

Cons

Built for convenience, not throughput. For many concurrent users vLLM is the better server, and running open weight models means you now own model version control and evaluation too.

#6★ 92,645· Python· last push 2026-09-25

vLLM

vllm-project/vllm · Apache 2.0

The high-throughput inference engine behind most serious self-hosted LLM deployments.

vLLM serves open weight models at production scale using PagedAttention and continuous batching, and exposes an OpenAI compatible API so application code does not need to change. If your plan is to self-host a model for data residency or cost reasons, this is the engine you will most likely end up running.

Key features

  • PagedAttention and continuous batching for high throughput
  • OpenAI compatible server out of the box
  • Tensor and pipeline parallelism across GPUs
  • Broad support for quantisation formats and hardware

Try it

pip install vllm

Controls it helps evidence

ISO 27001 A.8.6 capacity management and A.8.9 configuration management, ISO 42001 AI system operation

Honest take

Genuinely production grade, and the operational burden is the real cost. Self-hosting shifts GPU capacity, patching and model provenance from a vendor’s SOC 2 report onto yours, so budget for that before celebrating the lower token price.

Cons

GPU first and heavy to operate. Configuration is deep, and a fast moving release cadence means you need a real upgrade and regression process.

#7★ 59,579· Python· last push 2026-09-25

LiteLLM

BerriAI/litellm · MIT core, enterprise directory separately licensed

An AI gateway that puts 100+ model providers behind one OpenAI style API.

LiteLLM is both a Python SDK and a proxy server. Run the proxy and every team calls one endpoint, with virtual keys, per team budgets, spend tracking, logging and guardrail hooks applied centrally. It turns scattered provider keys into one controlled egress point, which is the fix for most of the shadow AI problems we write about.

Key features

  • One API across OpenAI, Anthropic, Bedrock, Vertex, Azure and self-hosted models
  • Virtual keys with budgets and rate limits per team or project
  • Spend tracking and request logging to your own database
  • Fallbacks, retries and load balancing across deployments

Try it

uv tool install 'litellm[proxy]'

Controls it helps evidence

ISO 27001 A.5.15 access control, A.8.15 logging and A.8.16 monitoring, SOC 2 CC6.1 and CC7.2, ISO 42001 AI system inventory

Honest take

A gateway is the single most useful AI control you can deploy, and LiteLLM is the most complete open source one. It is also a reminder of why the next section exists: in March 2026 attackers used a compromised Trivy GitHub Action in LiteLLM’s CI to steal its PyPI token and publish backdoored versions 1.82.7 and 1.82.8, and the proxy itself has had several security advisories this year, including SSRF and an MCP authentication bypass. Pin versions, verify hashes, patch quickly and treat the component that holds every provider key as your most sensitive dependency.

Cons

SSO, audit logs and some admin features sit in the enterprise tier. It becomes a critical path service, so it needs the same availability and patching discipline as your API gateway.

#8★ 35,026· TypeScript· last push 2026-09-24

Langfuse

langfuse/langfuse · MIT core, ee directories separately licensed

Open source tracing, evals and prompt management for LLM applications.

Langfuse records every model call, tool call and agent step as a trace, with cost, latency and the exact prompt and output. On top of that it runs evaluations and manages prompt versions. It is now owned by ClickHouse, and it remains self-hostable, which matters when the traces contain customer data.

Key features

  • Traces for LLM calls, tools and multi step agents
  • Evaluations with LLM as judge, human annotation and datasets
  • Prompt management with versioning and rollout
  • SDKs for Python and TypeScript plus OpenTelemetry support

Try it

git clone https://github.com/langfuse/langfuse.git && cd langfuse && docker compose up

Controls it helps evidence

ISO 27001 A.8.15 logging, SOC 2 CC7.2, EU AI Act Article 12 record keeping, ISO 42001 AI system monitoring

Honest take

If a customer or regulator asks what your AI feature said to someone on a given day, this is how you answer. Decide your trace retention and redaction rules on day one, because a tracing tool that stores full prompts is also a new store of personal data.

Cons

Self-hosting v3 means running Postgres, ClickHouse, Redis and blob storage, which is more than a quick docker compose suggests for production. Some features sit under the enterprise licence.

Test and guard AI applications

Red teaming, vulnerability scanning, runtime guardrails and PII handling for LLM features, the layer most teams still test by hand.

#9★ 25,428· TypeScript· last push 2026-09-25

Promptfoo

promptfoo/promptfoo · MIT

Evals and automated red teaming for prompts, agents and RAG, run from the command line or CI.

Promptfoo lets you describe test cases in a YAML file and run them against any model, then generates adversarial attacks such as prompt injection, jailbreaks and data exfiltration against your own application. OpenAI agreed to acquire Promptfoo in March 2026 and committed to keeping the open source tool under its current licence.

Key features

  • Declarative test suites that compare models and prompts side by side
  • Red team plugins mapped to OWASP LLM Top 10 and other frameworks
  • CI integration that fails a build on regressions
  • Runs locally, so test data does not need to leave your machine

Try it

npx promptfoo@latest init

Controls it helps evidence

ISO 27001 A.8.29 security testing, SOC 2 CC7.1 and CC8.1, ISO 42001 AI system verification and validation, EU AI Act Article 15

Honest take

The quickest way to turn we tested the chatbot into a repeatable artefact with a date and a pass rate. Wire the red team run into CI so every prompt or model change produces evidence, rather than running it once before launch.

Cons

Now owned by a model vendor, which some buyers will want to note in a supplier review even with the licence commitment. Generated attacks need human triage to separate real issues from noise.

#10★ 9,350· Python· last push 2026-09-16

garak

NVIDIA/garak · Apache 2.0

NVIDIA’s LLM vulnerability scanner, a probe library for model level weaknesses.

garak works like a network vulnerability scanner for language models. It fires a large library of probes at a model, covering prompt injection, encoding tricks, jailbreaks, data leakage and toxicity, and uses detectors to decide whether each attempt landed. Where Promptfoo is strongest on your application, garak is strongest on the model underneath it.

Key features

  • Large probe library across injection, jailbreak, leakage and misuse
  • Targets for OpenAI, Hugging Face, Bedrock, REST endpoints and more
  • Detectors that score each probe automatically
  • Reports in JSONL for tracking results over time

Try it

python -m pip install -U garak

Controls it helps evidence

ISO 27001 A.8.29 security testing, ISO 42001 AI system verification and validation, EU AI Act Article 15 robustness

Honest take

Most useful when you are choosing or self-hosting a model and need evidence that you compared candidates on something other than benchmarks. Run it against a new model before it goes into your gateway, and keep the report with the approval.

Cons

A full run is slow and can be expensive against paid APIs. Results describe the model, not your application, so pair it with application level testing.

#11★ 7,188· Python· last push 2026-09-24

NeMo Guardrails

NVIDIA-NeMo/Guardrails · Apache 2.0

Programmable input, output, dialog and retrieval rails for LLM applications.

NeMo Guardrails sits between your application and the model and enforces rules you define: block certain topics, check inputs for jailbreak attempts, fact check outputs against retrieved sources, or require a tool call before an answer. Rails are written in Colang, a small language for describing conversation flows, plus Python actions.

Key features

  • Input, output, dialog, retrieval and execution rails
  • Built in jailbreak, moderation and hallucination checks
  • Colang for describing allowed conversation flows
  • Integrates with LangChain and most model providers

Try it

pip install nemoguardrails

Controls it helps evidence

ISO 42001 Annex A controls for AI system operation, EU AI Act Article 14 and 15, ISO 27001 A.8.28

Honest take

The right tool when a policy needs to be enforced in code rather than hoped for in a system prompt. Keep the rail configuration in the same repo and review process as the application, because a rail change is a behaviour change.

Cons

Colang has a learning curve, and each rail that calls a model adds latency and cost. Rails reduce risk, they do not remove it, so keep testing with the tools above.

#12★ 11,028· Python· last push 2026-09-24

Presidio

data-privacy-stack/presidio · MIT

Detect, redact and anonymise personal data in text, images and structured data.

Presidio finds PII such as names, emails, phone numbers, card numbers and national identifiers using a mix of NLP models, patterns and checksums, then masks, replaces or encrypts it. It started life at Microsoft as microsoft/presidio and now lives under the data-privacy-stack organisation. For AI teams it is the standard way to scrub prompts, logs and training data before they reach a model or a trace store.

Key features

  • Analyzer for PII detection with configurable recognisers
  • Anonymizer for masking, replacement, hashing and encryption
  • Image redaction, including DICOM medical images
  • Custom recognisers for your own identifiers

Try it

pip install presidio_analyzer presidio_anonymizer && python -m spacy download en_core_web_lg

Controls it helps evidence

ISO 27001 A.5.34 PII protection and A.8.11 data masking, PCI DSS Requirement 3, ISO 42001 data for AI systems, GDPR data minimisation

Honest take

The component you add in front of a model when your data processing agreement says personal data should not go to the provider. Put it in your gateway or trace pipeline, not in each application, so there is one place to tune and evidence it.

Cons

Detection is probabilistic and needs tuning per language and data type, especially for Australian identifiers such as TFNs and Medicare numbers. Treat it as risk reduction, not a guarantee.

Secure the software supply chain

Secrets, dependencies, code and CI pipelines. The unglamorous controls that stopped being optional after this year of package compromises.

#13★ 29,472· Go· last push 2026-09-23

Gitleaks

gitleaks/gitleaks · MIT

Fast secret scanning for git history, directories and stdin.

Gitleaks scans commits and files for API keys, tokens and credentials using a large set of regex and entropy rules, and runs as a pre-commit hook, a CI step or a one off audit of a repository’s full history. With AI provider keys now worth real money to attackers, a leaked key is a billing incident as well as a security one.

Key features

  • Scans full git history, a directory or piped input
  • Hundreds of built in rules plus custom TOML rules
  • Baselines so you can adopt it on an old repo without noise
  • Pre-commit hook and GitHub Action

Try it

brew install gitleaks && gitleaks git -v

Controls it helps evidence

ISO 27001 A.5.17 authentication information and A.8.4 access to source code, SOC 2 CC6.1, PCI DSS Requirement 8

Honest take

The first tool on this list we would install, because it takes five minutes and finds something in almost every older repo. Run it on history once, rotate what it finds, then keep it in pre-commit so it stays clean.

Cons

It tells you a string looks like a secret, not that the secret is live. TruffleHog verifies credentials against the provider, which is useful for triage, under the more restrictive AGPL licence.

#14★ 38,068· Go· last push 2026-09-24

Trivy

aquasecurity/trivy · Apache 2.0

One scanner for vulnerabilities, misconfigurations, secrets and SBOMs across images, code and clusters.

Trivy scans container images, filesystems, repositories, Kubernetes clusters and infrastructure as code, and can generate or consume SBOMs in CycloneDX and SPDX. For a small team it replaces three or four separate tools with one binary and one report format.

Key features

  • Vulnerability scanning for OS packages and language dependencies
  • Misconfiguration checks for Terraform, Kubernetes and Dockerfiles
  • Secret scanning and licence detection
  • SBOM generation for EU Cyber Resilience Act and customer requests

Try it

brew install trivy && trivy fs --scanners vuln,secret,misconfig .

Controls it helps evidence

ISO 27001 A.8.8 technical vulnerabilities, SOC 2 CC7.1, PCI DSS Requirements 6.3 and 11.3, EU Cyber Resilience Act SBOM expectations

Honest take

Still the best all in one open source scanner. It is also the clearest lesson of 2026. In March attackers published a malicious Trivy 0.69.4 release and force-pushed 76 of the 77 version tags in trivy-action, plus every tag in setup-trivy, to point at credential stealing code (CVE-2026-33634). Anyone referencing those actions by tag ran the attacker’s code, and the secrets it stole led to the backdoored LiteLLM releases. Workflows pinned to a full commit SHA were not affected. Pin every action that way, and scope CI tokens so a scanner cannot publish your packages.

Cons

Vulnerability output is noisy without triage rules. The deeper cloud and runtime features are in Aqua’s commercial platform.

#15★ 16,756· C· last push 2026-09-25

Semgrep

semgrep/semgrep · LGPL 2.1 engine, rules under a separate licence

Fast static analysis with rules that look like the code they match.

Semgrep finds bugs and insecure patterns across 30+ languages using rules that read like source code, which means security engineers and developers can both write them. It is well suited to catching the specific mistakes your own codebase keeps making, including ones introduced by coding agents.

Key features

  • Pattern based rules that are easy to read and write
  • Large public rule registry, including OWASP Top 10 coverage
  • Fast enough to run on every pull request
  • Autofix suggestions for many rules

Try it

brew install semgrep && semgrep scan --config auto

Controls it helps evidence

ISO 27001 A.8.28 secure coding and A.8.29 security testing, SOC 2 CC8.1, PCI DSS Requirement 6.2

Honest take

The best way to encode we never do X here as a check that runs on every change, which matters more now that a coding agent writes a lot of the code. Start with the default rules, then add a handful of custom rules for your own footguns.

Cons

Cross file and cross function analysis are in the paid tier. Semgrep changed the licence on its maintained rules in late 2024, which led to the community Opengrep fork, so check what you are allowed to do with the rules before redistributing them.

#16★ 6,574· Rust· last push 2026-09-24

zizmor

zizmorcore/zizmor · MIT

Static analysis for GitHub Actions workflows.

zizmor audits your .github/workflows files for the mistakes attackers actually exploit: template injection from untrusted input, over broad token permissions, unpinned third party actions, cache poisoning and credentials persisted into artifacts. After a year in which compromised actions were the entry point for several major package attacks, it is the fastest way to find out how exposed your pipelines are.

Key features

  • Template injection and dangerous trigger detection
  • Excessive permissions and credential persistence checks
  • Unpinned and impostor action detection
  • SARIF output for GitHub code scanning

Try it

brew install zizmor && zizmor .

Controls it helps evidence

ISO 27001 A.8.25 secure development lifecycle and A.8.32 change management, SOC 2 CC8.1, PCI DSS Requirement 6

Honest take

Run it once on your busiest repo and you will almost certainly find an action pinned to a tag and a workflow with write permissions it does not need. It pairs directly with the Trivy lesson above: pinning is the control, and zizmor tells you where it is missing.

Cons

GitHub Actions only. It will not catch a pinned SHA that points at malicious code, so pinning still needs a review step.

Compliance and cloud posture as code

Open source scanners and policy engines that map findings to SOC 2, ISO 27001 and PCI DSS, and produce evidence an auditor can re-run.

#17★ 14,863· Python· last push 2026-09-24

Prowler

prowler-cloud/prowler · Apache 2.0

Open source cloud security posture checks mapped to SOC 2, ISO 27001, PCI DSS and more.

Prowler runs hundreds of checks against AWS, Azure, Google Cloud, Kubernetes, Microsoft 365, GitHub and other providers, and maps each finding to compliance frameworks. On AWS alone it ships 662 checks and 50 frameworks, including SOC 2, ISO 27001:2022 and PCI DSS 4.0, which makes it the closest open source equivalent to the cloud tests inside Vanta or Drata.

Key features

  • CLI, API and web UI
  • Compliance reports per framework with pass and fail per control
  • Checks across major clouds, Kubernetes, M365 and GitHub
  • Outputs to CSV, JSON, OCSF and AWS Security Hub

Try it

pip install prowler && prowler aws --compliance soc2_aws

Controls it helps evidence

SOC 2 CC6 and CC7, ISO 27001 A.5.23 cloud services and A.8.9 configuration management, PCI DSS Requirements 1, 2 and 10

Honest take

If you are preparing for a first SOC 2 or ISO 27001 audit on AWS, run this before you buy anything. The report tells you how far you are from a clean cloud baseline, and a dated, re-runnable scan is good evidence in its own right.

Cons

A posture scanner, not a compliance programme: it does not do policies, vendor reviews, access reviews or auditor workflow. Expect a long first report that needs triage.

#18★ 9,028· Python· last push 2026-09-24

Checkov

bridgecrewio/checkov · Apache 2.0

Policy as code scanning for Terraform, CloudFormation, Kubernetes, Helm and more.

Checkov scans infrastructure as code before it is deployed and flags misconfigurations such as public buckets, unencrypted storage and over permissive security groups. It shifts the same checks Prowler runs against live cloud accounts left into the pull request, where fixing them is cheap.

Key features

  • Over a thousand built in policies across major clouds
  • Terraform plan scanning, not just static files
  • Custom policies in Python or YAML
  • Inline suppressions with a written justification

Try it

pip install checkov && checkov -d .

Controls it helps evidence

ISO 27001 A.8.9 configuration management and A.8.27 secure architecture, SOC 2 CC8.1, PCI DSS Requirements 1 and 2

Honest take

Most useful when you treat suppressions as the real output. Every skipped check should carry a reason in the code, which gives an auditor a clear record of accepted risk instead of a wall of red.

Cons

Noisy by default, and teams often give up on it rather than baselining. Owned by Palo Alto Networks through the Bridgecrew acquisition, so some features point toward Prisma Cloud.

#19★ 12,270· Go· last push 2026-09-24

Open Policy Agent

open-policy-agent/opa · Apache 2.0

A general purpose policy engine for authorisation, admission control and CI gates.

OPA evaluates policies written in Rego against any JSON input, which lets you apply the same rules to Kubernetes admission, API authorisation, Terraform plans and CI pipelines. It is a CNCF graduated project and the usual answer when an organisation wants one policy language instead of a rule engine per tool.

Key features

  • Rego policy language with a test framework
  • Runs as a library, sidecar, daemon or CLI
  • Gatekeeper for Kubernetes admission control
  • Decision logs for audit trails

Try it

brew install opa && opa eval --data policy.rego --input input.json 'data.example.allow'

Controls it helps evidence

ISO 27001 A.5.15 access control and A.8.9 configuration management, SOC 2 CC6.1 and CC8.1, ISO 42001 AI system operation

Honest take

Worth the learning curve once you have more than a couple of places where the same rule should apply. Decision logs are the underrated feature: they record every allow and deny with its input, which is exactly the evidence an access control audit asks for.

Cons

Rego is unfamiliar to most developers and takes time to learn well. For simple cases, a tool’s built in rules are faster to live with.

#20★ 16,992· C++· last push 2026-09-25

Wazuh

wazuh/wazuh · GPL 2.0

Open source SIEM and XDR with file integrity, configuration assessment and compliance mapping.

Wazuh collects logs and security events from endpoints, servers and cloud workloads, then adds file integrity monitoring, vulnerability detection and configuration assessment against CIS benchmarks. Alerts come mapped to PCI DSS, GDPR, HIPAA, NIST 800-53 and the SOC 2 trust services criteria, which is why it shows up in so many first time compliance programmes.

Key features

  • Agents for Windows, macOS and Linux plus cloud log ingestion
  • File integrity monitoring and configuration assessment
  • Vulnerability detection across the fleet
  • Dashboards and alerts mapped to compliance frameworks

Try it

curl -sO https://packages.wazuh.com/4.14/wazuh-install.sh && sudo bash ./wazuh-install.sh -a

Controls it helps evidence

ISO 27001 A.8.15 logging and A.8.16 monitoring, SOC 2 CC7.2, PCI DSS Requirements 10 and 11.5

Honest take

The most capable free answer to how do you monitor for security events, and it covers several PCI DSS requirements that small teams otherwise pay a lot to meet. It is a platform to run, not a tool to install, so assign an owner and a weekly alert review from the start.

Cons

Heavy to operate at scale, with an indexer cluster to size and maintain. Out of the box alert volume is high and needs tuning before anyone will read it.

Dropped from the list, and why

If any of these are in your dependency inventory, they deserve a ticket.

  • continuedev/continue: The README now says the repository is no longer actively maintained and read-only. Version 2.0.0 was the final release, so it should not be a new dependency.
  • protectai/llm-guard: Now archived on GitHub, with its last commit in July 2026. It was a strong input and output scanner, and anyone still running it should plan a move to NeMo Guardrails or a gateway guardrail.
  • strongdm/comply: An early open source SOC 2 policy framework with no push since July 2022. Fine for reading, not for running a compliance programme in 2026.

Near misses worth knowing

  • trufflesecurity/trufflehog for verified secret detection
  • anchore/syft for SBOM generation
  • ggml-org/llama.cpp, the engine under many local runtimes
  • open-webui/open-webui for a self-hosted chat interface
  • CISOfy/lynis for host hardening audits
  • usnistgov/OSCAL for machine readable control catalogues
  • ossf/scorecard for scoring your dependencies
  • sigstore/cosign for signing images and binaries

Frequently asked questions

How did you choose these repositories?
Each repo had to be actively maintained on the day we checked, useful to a team shipping AI features in a regulated or security conscious business, and relevant to at least one control in SOC 2, ISO 27001, ISO 42001 or PCI DSS. We read every README and checked every install command against it. Popularity was not enough on its own, which is why some highly starred repos are in the near misses.
Are the star counts accurate?
They come from the GitHub API on the date shown at the top of the page, and we refresh them with a script rather than by hand. Stars measure attention, not quality or security, so we never rank on them.
Can open source tools replace Vanta, Drata or another compliance platform?
They can cover a lot of the technical testing. Prowler, Checkov, Trivy and Wazuh between them produce much of the cloud, code and monitoring evidence a SOC 2 or ISO 27001 audit asks for. What they do not do is policy management, access reviews, vendor risk, training tracking and the auditor workflow, which is most of what the platforms charge for. Our compliance software comparison covers that trade-off.
Is it safe to run AI coding agents like Claude Code or Codex on company code?
Yes, with the same care you would give any tool that can run commands with your permissions. Use the sandbox or permission modes, commit shared settings, treat plugins and MCP servers as third party code, and keep secrets out of the working directory. Our blog covers the specific incidents from 2026 that make each of those steps worth doing.
Which repo should a small team install first?
Gitleaks on your main repositories, then zizmor on your GitHub Actions workflows. Both take minutes, both almost always find something real, and neither needs any infrastructure.
What is the most useful single control for AI risk?
A gateway such as LiteLLM in front of every model provider, with tracing in Langfuse behind it. Once all model traffic goes through one place you can apply budgets, redaction and guardrails centrally, you have logs that answer what the model said, and any other traffic to a provider API stands out as something to investigate.
Do open source security tools create their own supply chain risk?
Yes, and 2026 made that concrete. Attackers hijacked the version tags on the Trivy GitHub Actions and used them to steal the credentials that published backdoored LiteLLM releases. Projects that pinned the actions to a commit SHA were not affected. Pin actions and packages to verified hashes, give CI jobs the least permission they need, and include your security tooling in your dependency inventory.
Does the open source licence matter for a commercial product?
It does. MIT and Apache 2.0 are permissive. GPL and AGPL, used by Wazuh and TruffleHog, carry obligations if you distribute or offer the software as a service. n8n uses a fair-code licence that restricts offering n8n itself as a service, and Claude Code is not open source at all. Check each one before it becomes part of what you sell.

Turn the tools into evidence

// Signal, not noise

Get the AI tools shortlist

The tools worth paying for, the deals worth taking, a short, no-spam email.