Denali is an independent, open-source AI security platform for evidence-backed inventory, posture, attack paths, and runtime context. It keeps observed facts, inferred relationships, security conclusions, and coverage limits separate so an evaluator can see both what Denali knows and what it could not verify.
The prioritized product direction and release gates are recorded in the agent security roadmap.
The repository contains a runnable local product: a PostgreSQL-backed API, a standalone web application, deterministic issue and detection engines, provider onboarding, first-party collectors, and bounded importers. It is not an extension of a CSPM. Prowler is supported as an external findings source through OCSF.
Status terms in this README are deliberately independent:
- Shipped means the capability is implemented on
mainand covered by repository tests and/or a production web or container build. - Locally accepted means the customer-like workflow was also completed against live provider resources from a local Denali deployment.
- Pending live acceptance means the implementation and local verification exist, but the documented human, browser, or provider end-to-end pass is incomplete.
- Planned means the capability is not implemented in this repository.
| Capability | Implementation | Acceptance evidence |
|---|---|---|
| Evidence store, API, web UI, connectors, importers, issues, and runtime detections | Shipped | Locally verified by the automated and PostgreSQL suites and the production web build |
| Self-service AWS connection | Shipped | Locally accepted against a live AWS account, including Quick Create, exact account binding, enabled-Region discovery, and independent plane validation |
| Self-service Microsoft Azure connection | Shipped | Pending live acceptance. Live subscription setup and healthy validation succeeded after RBAC propagation, but the final connected-browser, unselected-subscription, partial-state, disable, and delete pass remains open |
| Self-service Google Cloud connection | Shipped | Locally accepted against three live projects, with a unique keyless principal and all declared project/plane validations passing |
| Self-service Microsoft Entra connection | Shipped | Pending live acceptance. Production admin consent and validation were observed, but a complete retained lifecycle and collection record remains open |
| Self-service Google Workspace connection | Shipped | Pending live acceptance. Fixed-scope domain-wide delegation, validation, durable collection, and tenant isolation have automated coverage; the first hosted lifecycle and collection record remains open |
| Self-service GitHub connection | Shipped | Locally accepted through the organization-owned GitHub App against 18 exact repositories, with all 54 repository/plane validations passing |
| Self-service Azure Repos connection | Shipped | Hosted accepted on 10 September 2026 for lifecycle and code-to-cloud behavior; a dedicated dated record using the complete P0 evidence template remains to be retained |
| GCP and Azure code-to-cloud correlation | Shipped | Locally accepted against independently observed, private scale-to-zero fixtures with exact source identity, PostgreSQL reporting, and browser evidence |
| AWS Lambda, ECS, EKS, and SageMaker code-to-cloud correlation | Shipped | Locally accepted against exact live account/Region validation and eight independent deployment collection planes |
| AWS AgentCore runtime detection and response | Shipped | Pending live acceptance. Metadata-only CloudWatch span collection, five-minute durable polling, exact session correlation, drift/tool/sequence detections, and manual maker-checker response approval pass automated and PostgreSQL verification; hosted AWS create/update, collect, investigate, and approve remains required |
| Azure Foundry runtime detection and response | Shipped | Production accepted on 14 September 2026 against the Anna reference agent: one durable metadata-only session, six agent/model/tool activities, complete Application Insights coverage, and zero prohibited content fields |
| Shared EKS, GKE, and AKS workload correlation | Shipped | Locally accepted through a live, control-plane-only EKS fixture with exact workload UID/revision, service-account, image-digest, negative-case, persistence, API, and teardown evidence; GKE and AKS workload identities remain covered by automated contract tests |
| GitHub source-to-cloud correlation | Shipped | Locally accepted against two immutable GitHub revisions, with all source/inventory/posture/correlation planes complete and two independently observed runtime links proven |
| Two-application code-to-cloud Golden Path | Shipped | Anna on AWS and Summit on GCP are bounded by a versioned reset/verify manifest with exact source-to-runtime links, declared model/tool/action context, real Vertex activity, Entra discovery context, three image vulnerabilities, two correlated governance issues, and a dedicated dashboard story |
These connection statuses describe onboarding and access validation. They do not mean that provider collection ran or that inventory, findings, or a security verdict exist.
Denali presents the following product surfaces in the web application and API:
- AI inventory and evidence-backed relationships for agents, applications, models, MCP servers and tools, guardrails, frameworks, pipelines, data stores, workloads, repositories, identities, and software components.
- Atomic configuration and imported findings, scanner-neutral vulnerabilities, and deterministic issues assembled only from sufficient independent evidence.
- GitHub-backed, immutable-revision source collection and code-to-cloud views that require exact deployment identifiers, preserve unmatched and ambiguous candidates, and keep artifact identity separate from unattested source revision claims.
- Provider-neutral runtime activity plus deterministic, evidence-linked runtime detections, ordered AWS AgentCore and Azure Foundry session investigations, and approval-gated manual response requests.
- Source coverage that keeps complete, partial, failed, unsupported, and unknown states visible.
- Stable application routes with direct deep links and browser Back/Forward navigation.
- A manifest-bounded Golden Path dashboard for a small, reproducible two-application demo; see the operator and presenter guide.
- Connection lifecycle and access validation for AWS, Azure, Google Cloud, and GitHub, including explicit scope selection and disable-before-delete safeguards.
Implemented collection and import paths include:
| Source | Current path |
|---|---|
| Source repositories | Local or GitHub App-backed static Python, TypeScript, JavaScript, Terraform, SAM/CloudFormation YAML and JSON, Cloud Run YAML, Kubernetes YAML, ARM JSON, and Bicep analysis; bounded repository posture; exact-identifier code-to-cloud correlation |
| MCP Streamable HTTP | Initialization and paginated tools/list observation without tool invocation |
| AWS | Bedrock Agents Classic, AgentCore, bounded Lambda/ECS/EKS/SageMaker deployment inventory, CloudFormation-stack inventory and posture, Bedrock management activity from CloudTrail Event History, and metadata-only AgentCore OpenTelemetry/OpenInference spans from CloudWatch Logs |
| Google Cloud | Cloud Run, Cloud Run functions Gen2, and GKE cluster inventory through Cloud Asset RESOURCE snapshots; Vertex AI audit activity from Cloud Logging |
| Microsoft Azure | Container Apps, Function Apps, and AKS cluster inventory through Azure Resource Graph with exact Azure code-to-cloud identity; opt-in, metadata-only Foundry agent/model/tool spans through Application Insights; Entra activity remains a separate connector |
| Microsoft Entra | AI application, permission, sign-in, and application-management collection through a separate Microsoft Graph connector |
| External findings and scanners | OCSF findings, Syft SBOMs, and Grype vulnerability reports |
| Runtime exports | AWS Bedrock CloudTrail, Google Cloud Vertex AI, Google Workspace Gemini, and Microsoft Entra AI sign-in JSON; hosted AgentCore and Azure Foundry spans use their existing keyless or consent-based cloud connections instead of browser upload |
Provider validation and collection remain separate boundaries. A healthy connection does not claim that collection ran. GitHub validation reads no source blobs; an explicit collection action separately resolves each selected repository to an immutable revision, applies hard tree/file/byte limits, analyzes a temporary snapshot, and discards its token and source files.
Denali's implementation follows a few non-negotiable rules:
- Connection health proves only that declared, read-only entrypoints were callable for the recorded identity and scope at a recorded time.
- Collection coverage is separate from connection health. A failed or partial collection cannot withdraw previously observed inventory or resolve findings by absence.
- Findings and runtime events do not manufacture inventory assets or graph edges. Identity and correlation require exact, independently observed identifiers.
- Inventory is not a finding; a finding is not an issue; activity is not a detection; and none of these is automatically a risk verdict or confirmed incident.
- Unsupported, unselected, unavailable, and unknown scope remains visible rather than becoming an empty or safe result.
- Current collectors retain bounded metadata and evidence. Customer cloud credentials and transient provider access tokens are not written to evidence or connection records; raw one-time setup capabilities are not retained after use. Raw prompts or responses, secret values, and arbitrary scanner payloads are also excluded.
- Demo records are visibly identified as fixture evidence.
The canonical model is richer than OCSF because it must preserve durable asset identity and relationships. OCSF remains an interchange boundary for findings and activity, not a source of invented Denali graph identity.
Docker is the shortest path to the product:
docker compose up -d --build
docker compose exec api denali-demo-seedOpen the web application at http://127.0.0.1:3080. The API is at
http://127.0.0.1:8088, with interactive documentation at
http://127.0.0.1:8088/docs. PostgreSQL is exposed on 127.0.0.1:55450.
The seed command creates clearly labelled fixture inventory, findings, vulnerabilities, issues, runtime activity, and detections so the evaluator can inspect the evidence and coverage UX without provider credentials. A useful first pass is:
- Review Inventory and open a resource's assertions and relationships.
- Compare Findings, Vulnerabilities, and Issues to see atomic facts kept separate from composed attack paths.
- Compare Runtime activity with Runtime detections.
- Review Sources for explicit coverage state.
- Open Connections to inspect the shipped onboarding flows. Actions that depend on an external operator identity or artifact publisher remain unavailable until configured.
The local Compose runtime uses one configured tenant and has no end-user authentication or
authorization layer. It is an evaluation and development stack, not a production deployment
recipe. Data persists in the denali-postgres Docker volume.
For the Clerk Organizations, Vercel, Modal, and Neon pilot deployment, start with the ordered launch checklist, then use the hosted multi-tenant pilot runbook. Hosted mode maps each active Clerk organization to a stable Denali tenant UUID, enforces admin-only mutations, and runs provider validation through durable Modal jobs. The canonical service boundaries, secret ownership, tenant isolation rules, and durable-work standard are recorded in ADR 0028.
Optional PostHog browser analytics records anonymous page views and masked clicks. Set the
public VITE_POSTHOG_PROJECT_TOKEN and VITE_POSTHOG_HOST build variables only for the
environments intended to collect usage; an empty token disables capture. See the
product analytics runbook for the privacy boundary,
configuration, and verification steps.
For CLI work, install the provider extras you need and point the commands at the local database:
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -e '.[api,aws,azure,gcp,github,dev]'
export DENALI_DSN=postgresql://denali:denali-local@127.0.0.1:55450/denaliThe installed entrypoints include denali-repo-scan, denali-mcp-observe,
denali-agentcore-scan, denali-aws-scan, denali-aws-stack-scan,
denali-aws-stack-posture, denali-repository-posture, denali-code-to-cloud,
denali-kubernetes-import, denali-ocsf-import, denali-syft-import, denali-grype-import,
denali-activity-import, denali-aws-runtime, denali-gcp-vertex-runtime,
denali-gcp-deployments-scan, and
denali-entra-scan. Use each command's --help output for its exact scope and inputs.
Self-service onboarding requires operator-owned integration identities and, for AWS, Azure, and GCP setup artifacts, a private short-lived publisher. The customer grants only the documented read scope and retains control of provider-side IAM or App installation state. Denali does not accept customer access keys, service-account JSON keys, personal access tokens, or cloud CLI sessions.
The accepted contracts, operational requirements, validation planes, lifecycle behavior, and evidence limits are documented here:
- ADR 0018 — AWS role onboarding
- ADR 0019 — Azure multi-tenant application and selected subscriptions
- ADR 0020 — GCP keyless principal and selected projects
- ADR 0021 — GitHub App and exact repository boundaries
- ADR 0022 — GitHub immutable source collection
- ADR 0029 — Microsoft Entra admin-consent onboarding
- ADR 0033 — Google Workspace domain-wide delegation
- ADR 0034 — Azure Repos exact-repository onboarding
- GitHub App registration and operator configuration
The latest acceptance records are the AWS, Azure, GCP, and GitHub checkpoints.
Start with the product and evidence boundaries, then follow only the slice being evaluated:
- Codebase and system map for the repository layout, end-to-end data flow, local and hosted topology, persistence, frontend, connectors, evaluation engines, security boundaries, tests, and safe change recipes
- Standalone product and OCSF boundary
- Hosted multi-tenant runtime for the Vercel, Clerk, Modal, and Neon deployment, tenant authorization, secrets, and durable jobs
- Shared GitHub pilot for reusing a Platform installation with exact repository boundaries and unchanged native connections.
- Shared Google Cloud pilot for keyless Platform onboarding, Denali reuse, compatibility and hosted release acceptance
- Evidence-bearing issues
- SBOM-first vulnerability model, artifact correlation, and package occurrence identity
- Evidence-led code to cloud, provider-neutral deployment identity, GCP code to cloud, Azure code to cloud, AWS deployment code to cloud, shared Kubernetes code to cloud, static artifact inclusion, and deployment artifact provenance
- Provider-neutral runtime activity, Entra AI application discovery and runtime, and runtime detections, including AWS AgentCore AIDR and Azure Foundry AIDR
Product-preview definitions remain available for inventory, configuration findings, and issues. The Microsoft slice is documented in AI application discovery.
Pending acceptance is not the same as planned implementation:
- Azure needs the remaining human browser, unselected-scope, partial-state, disable, and delete acceptance checks.
Documented planned or deferred capabilities are not shipped:
- Complete hosted create, setup/callback, validate, collect, disable, and delete acceptance for every enabled provider; the individual production records remain authoritative.
- Complete the shared P0 operating gate in the agent security roadmap, including production configuration enforcement, non-empty two-Organization isolation, split Neon roles, alerts/backups, and a restore drill.
- GitHub branch-protection and pull-request posture through a separate, explicitly granted Administration-read plane.
- GitHub installation/repository lifecycle reconciliation and GitHub Enterprise Server.
- Slack and Jira onboarding after GitHub acceptance.
- Application-wide typography and clearer elapsed-time context during bounded cloud-IAM propagation waits.
- Broader exact runtime-to-inventory correlation for provider-native agents, models, tools, and execution identities that currently remain unresolved references.
Prompt- and response-content telemetry is neither shipped nor silently planned into the current cloud roles. It remains an explicit product-policy decision that would require separate tenant opt-in, permissions, redaction, retention, encryption, and evidence semantics.
pytest
ruff check .
npm --prefix web ci
npm --prefix web run buildThe default test run skips PostgreSQL integration tests unless DENALI_TEST_DSN is set. Run
that contract gate explicitly against the local Compose database:
DENALI_TEST_DSN=postgresql://denali:denali-local@127.0.0.1:55450/denali \
pytest -q tests/test_inventory_postgres.pyApache License 2.0. See LICENSE.