Open-source AI Security Operations Center: alert fusion, LLM-agent triage, MITRE ATT&CK investigation, and a replayable decision ledger for every agent step. Self-hostable, runs with no API keys, MIT licensed. Ships an MCP server for Claude, Cursor and Continue.
Platform Threat Model (STRIDE)
Scope: the AiSOC platform as a whole, with the credential vault as the top asset. AiSOC's CredentialVault holds live credentials to CrowdStrike, AWS, Okta, Splunk, and GitHub. Compromising AiSOC compromises everything it connects to, so the platform's own security is a P0 product feature.
This document is Phase 1.6 of the world-class program. It is referenced from SECURITY.md. The agent/tool surface has its own model in agent-threat-model.md.
Assets (ranked)
- Connector credentials in the vault (
services/api/app/security/credential_vault.py) — the crown jewels. - The KEK /
AISOC_CREDENTIAL_KEYand any KMS key material. - Tenant data at rest (Postgres, ClickHouse, Neo4j, Qdrant, Redis, Kafka).
- The audit log / Investigation Ledger (must be tamper-evident).
- Platform admin + SSO/JWT signing secrets.
Trust boundaries
flowchart LR
Operator["Operator / SSO"] --> API["Core API"]
Tenant["Tenant users"] --> API
API --> Vault["CredentialVault"]
Vault --> KMS["KEK in KMS/HSM"]
Vault --> DB[("Postgres (wrapped DEK + ciphertext)")]
API --> Connectors["Connectors"]
Connectors -->|"least-privilege creds"| Vendors["CrowdStrike / AWS / Okta / Splunk / GitHub"]
STRIDE
Spoofing
- Threats: forged JWT/SSO to assume an operator/tenant; a connector impersonating another tenant.
- Controls: OIDC/SAML token issuance fails closed on placeholder secrets (
services/api/app/auth/); per-request tenant binding (app.current_tenant_id) + query-layertenant_idfilters; API keys are vault-encrypted.
Tampering
- Threats: editing ciphertext in the DB; altering the audit trail.
- Controls: authenticated encryption (Fernet HMAC / envelope DEK) makes tampered ciphertext fail closed; audit hash chain (
services/api/app/services/audit_hash.py) makes ledger edits detectable; audit rows reject UPDATE/DELETE by the app role (Phase 3.1 gate).
Repudiation
- Threats: an operator denies a credential change or an autonomous action.
- Controls: append-only audit log with actor + hash chain; ledger records the deciding model/tier.
Information disclosure (top risk for this asset)
- Threats: a DB dump or leaked
AISOC_CREDENTIAL_KEYexposes every connector credential; plaintext secrets in logs. - Controls:
- Envelope encryption (
services/api/app/security/envelope_cipher.py), opt-in viaAISOC_CREDENTIAL_ENVELOPE=local|aws. Each secret is encrypted under its own data-encryption key; only the wrapped DEK is stored beside the ciphertext. Withaws, the KEK never leaves KMS, so a database dump yields wrapped DEKs and ciphertext that are useless without KMSdecryptpermission. Per-secret DEKs also bound the blast radius: one compromised DEK exposes one secret, not the vault. Rotation is re-wrapping the DEKs (EnvelopeCipher.rewrap), which never touches plaintext. Written asvault:v2:. - Fernet authenticated encryption (
services/api/app/security/credential_vault.py) is the default and the read path for everything written before envelope mode was enabled (vault:v1:). Under v1 the risk a dump poses is bounded byAISOC_CREDENTIAL_KEYalone, which is why v2 exists. - Enabling v2 needs no migration — v1 tokens keep decrypting and rows upgrade the next time they are written. Disabling it after v2 rows exist fails closed rather than handing ciphertext to a connector as if it were a credential.
- Leaf-level encryption keeps structural fields queryable while every secret value is ciphertext.
- The vault refuses to boot without a key outside development; plaintext keys never log.
- Envelope encryption (
Denial of service
- Threats: flooding investigations to exhaust budget/compute (see agent model, Phase 1.5); KMS throttling.
- Controls: per-tenant cost governor (Phase 1.5); KMS calls are per-secret-write/rotation, not per-request, so KMS rate is bounded.
Elevation of privilege
- Threats: an over-scoped connector credential lets a vault compromise pivot into full cloud admin.
- Controls: least-privilege connector scopes (
connector-least-privilege.md) — document and audit the minimum IAM/API scope each connector needs; break-glass + rollback for autonomous actions (Phase 9).
Key rotation (tested)
Rotation is a re-wrap, not a re-encrypt of every secret body:
- Provision a new KEK in KMS (or add a new
LocalKeyManagerprimary, keeping the old as historical). - Run
EnvelopeCipher.rewrap(token)over every stored token: the DEK is unwrapped with its original KEK and re-wrapped with the current one. The ciphertext body is untouched and the plaintext never leaves KMS's reach. - Retire the old KEK once no token references it.
Gated by services/api/tests/test_envelope_cipher.py (round-trip, rotation + re-wrap, plaintext-never-in-token, fail-closed on tamper/wrong-KEK).
Residual risk
A live KMS decrypt permission on the running API host is still a compromise path (the app must be able to decrypt to use credentials). Mitigations: scope the KMS key policy to the API role only, enable KMS CloudTrail on decrypt, and alert on anomalous decrypt volume. HSM-backed KEKs raise the bar further.