Cynative

Cynative is an open-source security agent framework built by a foundation model lab focused on cybersecurity. The framework ships as a single binary CLI tool, licensed under Apache-2.0, and gives security engineers a way to run AI-powered agents against their cloud infrastructure with live, read-only access. The company is also developing a sovereign cybersecurity foundation model — currently in development — designed to run entirely within a team's own perimeter.

What It Is

Cynative is a CLI-based agent harness for security engineers. It connects to AWS, GCP, Azure, Kubernetes, GitHub, and GitLab as one unified system, then uses a frontier LLM of the user's choice to reason across that infrastructure and return verified security findings. The framework ships with 45 built-in agents covering privilege escalation, network exposure, data and storage exposure, credentials and secrets, supply chain and CI, detection and audit coverage, inference exposure, Kubernetes control plane, and hardening and resilience. Users can also write custom agents in a single markdown file using the same format as the built-ins.

How the Read-Only Safety Model Works

Unlike coding agents or MCP servers, Cynative enforces read-only access at three distinct layers before any credential is attached to a request:

• Action gate: Every operation is resolved to its required IAM actions using the cloud providers' own API definitions, then authorized against a read-only policy (SecurityAudit on AWS, roles/viewer on GCP, Reader on Azure). Anything classified as a write fails closed.
• Network pinning: Each request host is pinned to its mapped service and region, and the resolved IP is verified before connecting. The agent can only reach the infrastructure it is configured for.
• Credential scoping (AWS): For assumed-role identities, credentials are re-vended through STS scoped to SecurityAudit, so AWS IAM enforces the boundary independently of the action gate.

This architecture means the tool is safe to point at production environments. Secrets are redacted from tool output before being sent to the model, and every tool call is recorded to a fail-closed JSONL audit log.

Architecture and Code Execution

For bulk research tasks — checking every public S3 bucket, listing EKS clusters across all regions — Cynative writes and runs JavaScript in an ephemeral sandbox rather than issuing one tool call at a time. The sandbox exposes the agent's tools as async JavaScript functions, allowing the model to loop, filter, and fan out calls concurrently using a built-in mapConcurrent helper. Only what the script logs returns to the model, keeping token usage efficient. The sandbox has no network or host access of its own; it can only call the tools Cynative exposes.

LLM Provider Flexibility and Sovereign Deployment

Cynative connects to LLMs through the embedded Bifrost SDK and supports a wide range of providers out of the box, including OpenAI, Anthropic, Azure OpenAI, Amazon Bedrock, Google Vertex/Gemini, Cohere, Mistral, Groq, Ollama, and vLLM. This means teams can point the framework at a model already running inside their cloud account — so the model, the agent, and the data all share one account and nothing leaves the environment.

The company's longer-term product is a dedicated cybersecurity foundation model, currently in development, intended to run inside a team's own perimeter and be tunable on that team's own data. A design partner program is open for teams that want to shape what the model does.

Founders and Background

Both founders, Shaked Zin (CEO) and Yuri Shapira (CTO), came to Cynative from Palo Alto Networks, where they led engineering after it acquired their previous company, PureSec, which the About page describes as having pioneered cloud-native runtime and IAM security.

Update: v1.11.3

The latest release is v1.11.3, published on September 10, 2026. The repository was last updated on September 11, 2026, indicating active development. The project notes it deliberately collects no telemetry — the team states they do not know how many teams run it, by design.