National Cyber Warfare Foundation (NCWF)login

cynative for read-only agentic security audits across cloud, code and runtime


0 user ratings
2026-11-12 21:30:01
milo
Red Team (CNA)
"cynative

cynative is an open-source framework that runs LLM-driven security agents with live, read-only access to your AWS, GCP, Azure, GitHub and Kubernetes estate for authorized defensive auditing.








Toolcynative/cynative — open-source Go framework for building and running read-only security agents over cloud, code and runtime
CategoryAgentic security automation / cloud security auditing (cli-tool, devsecops)
Primary UseAuditing your own infrastructure for privilege escalation, public exposure, supply-chain risk and IaC drift via built-in agents like aws-privilege-escalation and k8s-pod-privilege
Safe UseDesigned for authorized security assessments of environments you own — the action-gate fails closed to a read-only IAM policy before any credential is attached
Telemetry NoteEvery tool call is written to a fail-closed JSONL audit log — if the log cannot be recorded the run aborts — and AWS sessions are STS-scoped, so defenders see a complete, tamper-evident trail of agent activity in CloudTrail and local logs

cynative positions itself as an open-source framework for security agents with live, read-only access to your infrastructure, and the framing is deliberate: rather than bolting a chat model onto a pile of MCP connectors, it runs frontier models across GitHub, GitLab, AWS, GCP, Azure and K8s as one logical system and returns verified answers. The invocation model is conversational — cynative "what in my cloud is publicly exposed that shouldn't be?" — but the execution model underneath is engineered: one question fans out across the stack, with generated research code running in an ephemeral sandbox that queries your APIs in parallel. Every finding is cross-checked and traced back to its origin, which is the property most ad-hoc LLM auditing setups conspicuously lack.


The headline number is 45 built-in agents embedded directly in the binary, each one a reviewed prompt targeting a specific question. The catalog splits as 20 AWS agents (aws-privilege-escalation, aws-public-storage, aws-unpatched-workloads, aws-supply-chain), 11 Azure agents (azure-keyvault-exposure, azure-storage-exposure), 5 GCP agents (gcp-public-bindings, gcp-static-credentials, gcp-inference-exposure), 4 GitHub agents (github-workflow-trust, github-unpatched-dependencies, github-branch-protection) and 5 Kubernetes agents (k8s-pod-privilege, k8s-self-managed-apiserver-access). What is notable is the transparency: cynative agents list shows every agent with its description, and cynative agents show prints the exact prompt file that would run. There is no hidden system prompt you have to reverse-engineer.


Extensibility is where the architecture gets interesting. An agent is just a markdown file: strict YAML frontmatter whose only key is description, followed by the prompt body, with the filename serving as the agent name. The documented workflow is to dump a built-in — cynative agents show aws-public-datastores > ~/.cynative/agents/my-aws-public-datastores.md — edit it, and run it with cynative -p --agent my-aws-public-datastores. User files in ~/.cynative/agents/ shadow built-ins of the same name, and agents list marks shadowed copies, so you always know which version is live. Because --agent composes with -p, --auto-approve, --config and piped stdin, the same file can run interactively while you develop it and non-interactively once it settles — a CI-friendly lifecycle.


The README's sharpest section is the comparison against a generic coding agent with MCP servers, and it reads like it was written by someone who has actually watched that pattern fail. The throughput argument is architectural: instead of one tool call per turn, cynative writes sandboxed code that fans out API calls concurrently, spending fewer tokens for faster answers. The findings argument is about verification — a cross-checking verifier confirms every finding against live evidence rather than emitting unverified model output. The rest of the table is about trust boundaries, and it is worth parsing line by line.


The read-only story is the core differentiator. It is not an opt-in filter; it fails closed. Every call is resolved to its required IAM actions and checked against a security-audit policy before a credential is attached, and the README gives a concrete example of why naive filters are insufficient: secretsmanager:GetSecretValue is an IAM Read action, so a read filter would allow it, while the SecurityAudit policy blocks it. Credentials are additionally scoped via an STS session restricted to read-only, meaning AWS itself enforces the boundary at the platform level, not just in client code. That dual enforcement — client-side gating plus server-side session policy — is the right shape for anything you point at production.


Secrets handling follows the same defensive posture. Secrets are redacted from tool output before anything is sent to the model, which addresses one of the most common objections to letting an LLM crawl an estate full of credentials. The sandbox runs generated research code with no host access and with the network pinned to your mapped services, containing the blast radius of whatever the model decides to write. And the audit trail is fail-closed: a JSONL log records every tool call, and if the tool cannot record, it aborts the run. For a defender reviewing this tool, that log is also the artifact you hand to a compliance team after the fact.


The sovereignty angle is more than marketing. cynative ships as a single static binary written in Go with connectors built in — no third-party MCP servers or skills running with your credentials — and it talks to LLMs through the embedded Bifrost SDK, supporting OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, Google Vertex/Gemini, Cohere, Mistral, Groq, Ollama, vLLM and more. Configuration is environment-driven (CYNATIVE_LLM_PROVIDER, CYNATIVE_LLM_MODEL), and the README suggests running it on an instance inside the very cloud it audits, through that cloud's managed inference, so nothing leaves your environment. For teams with data-residency constraints or a mandate against third-party SaaS, running Bedrock or a local Ollama endpoint closes the loop entirely.


Supply-chain hygiene is unusually thorough for a project of this age. Homebrew (brew install cynative/tap/cynative), Scoop for Windows, and an install script that verifies the download's SHA-256 against the release checksums.txt, failing closed. Releases ship a Sigstore keyless bundle (checksums.txt.sigstore.json) verifiable with cosign verify-blob, macOS packages are Developer ID signed, notarized and stapled, and every asset is covered by GitHub release attestations (gh release verify). The README even documents the limits honestly: cosign fetches Sigstore's trust root over the network unless you pass --trusted-root, and the signature proves origin and integrity but not which release a loose file set came from — you need the release URL or gh release verify for version binding.


Day-to-day usage patterns shown in the README read like a defender's notebook: cynative -p "which IAM roles can escalate to admin?", tracing high-risk cloud permissions back to the pull request that granted them, hunting cloud credentials leaked in source code and computing their current blast radius, and detecting drift — live cloud resources absent from IaC. The pipe-friendly design stands out: cat findings.json | cynative -p "triage these findings by exploitability" turns it into a triage layer over output from your existing scanner fleet rather than a replacement for them.


Where does this fit in an authorized workflow? The intended context is explicit: auditing infrastructure you own, with the action-gate and STS-scoped credentials providing structural guarantees that the agent cannot mutate state. The natural adopters are cloud security engineers and product-security teams drowning in AWS, Azure and GitHub findings who want an analyst that can hold the entire code-to-runtime graph in one query, and purple teams wanting reproducible, logged reconnaissance of their own estate. Watch the Apache-2.0 repo's trajectory — at 202 stars with OpenSSF best-practices badge already in place, the engineering discipline visible in the README suggests this is one to track rather than wait on.



Official project repository for cynative/cynative.

Download Tool

Educational analysis for authorized security professionals. Use only in controlled, authorized environments.






Source: OffensiveSec
Source Link: https://www.offsecblog.com/2026/11/cynative-for-read-only-agentic-security.html


Comments
new comment
Nobody has commented yet. Will you be the first?
 
Forum
Red Team (CNA)



Copyright 2012 through 2026 - National Cyber Warfare Foundation - All rights reserved worldwide.