Connector Security Profile
This profile fixes conservative deployment and operational defaults for draft ADRs0092-0096. Values below are platform design decisions, not limits prescribed by an RFC or a claim of compliance. They remain subject to implementation and the conformance gate.
Authority and rollout
Ship project-owned Connections first. Schema arms for person/workspace ownership remain policy-disabled until their offboarding/delegation gates pass. No transparent proxy, raw token issuance, node-scoped app or cross-project app borrowing is enabled implicitly. Explicit authenticated HTTPS relay is the ordinary provider traffic path; trusted MCP uses its own protocol/session grant.
Future Gmail onboarding uses a customer-supplied app first. External secret sources such as 1Password remain sources behind the existing secrets port, not an alternate credential execution or agent-token delivery plane. These are rollout defaults for deferred integrations, not claims of vendor approval.
The broker is a minimal module inside the secrets service and the only code performing token/registration calls with refresh tokens, client secrets or app private keys. Parsing ordinary provider responses and fetching discovery data stay outside that process. Use a sandboxed, keyless discovery worker with fixed CPU/memory/deadline budgets and host-enforced destination checks.
Credential-bearing relay/MCP workers serve one tenant per workload identity and process boundary. They share fleet infrastructure, not access-credential caches or wildcard custody authority. Their network role reaches approved provider origins and authorized custody operations only. They cannot mint grants, administer OpenBao or read another tenant's service subjects. A customer-managed key tenant follows the same boundary plus its key-veto requirements. Shared brokers still need per-tenant quotas and independently authenticated requests.
Workload authentication does not authorize a Connection. Custody independently checks the registered caller role, verified tenant, exact binding and audience, and live grant, as required by ADR#0095. Successful mTLS or a spiffe:// URI alone also does not prove that the runtime holding the key was attested.
The proposed production default is the attested workload identity profile, governed by draft ADR#0098. It prefers SPIRE and X.509-SVIDs with automatic credential and trust-bundle rotation. It must remain disabled until its implementation and deployment gates pass.
For a deployment claiming that profile, its selected issuer must verify trusted deployment evidence at bootstrap, issuance and renewal, and bind the verified runtime and workload identity to the requested public key. Possession of a signing key or a caller-supplied identity is insufficient. Missing, invalid or unavailable attestation refuses issuance or renewal without silently falling back to a static credential. This is an acceptance condition for that claim, not evidence of an implemented issuer. Workload attestation remains deferred in ADR#0053; ADR#0023's self-hosted bootstrap baseline remains a distinct compatibility option without implying attestation or satisfying the workload identity gate.
Provider authentication follows an explicit reviewed protection profile. Prefer supported workload federation or asymmetric client authentication and supported sender-constrained access tokens. Client mTLS alone does not sender-constrain a token; RFC8705, DPoP and AAuth use different proof contracts. A bearer-only provider retains an explicit compatibility profile and protected custody, never a claim that a stolen token cannot be replayed. Unsupported binding modes fail closed rather than downgrading silently.
Cache resource tokens only, isolated by verified tenant, exact Connection or federation binding, provider account/app, issuer, audience, granted scope, authorization/credential generations and holder key where applicable. Check live grant and policy authority on every use. Federation assertions are minted for one exchange audience and are never cached, persisted or exposed to agent code. Trust-material caches remain separate. The workload identity reference defines the corresponding rotation, cache, revocation and existing-session acceptance gates.
OAuth edge and discovery
Use one configured HTTPS console origin per environment for the callback and public Client ID Metadata Document. Configure an exact callback path; no user input, wildcard or arbitrary return URL chooses it. Missing origin configuration blocks app activation. Development has a distinct app and environment; a production callback never points to a development tunnel.
Use the reviewed MCP2026 authorization profile where applicable, independently of transport/session versions. Use authorization code with S256 PKCE. Discovery must explicitly support the required flow; absent evidence does not mean support. V1 grants no automatic provider exception for missing PKCE. The callback verifies the same initiating session, opaque initiator binding, browser binding, state, issuer, one-use claim and deadline. A public CIMD document contains no tenant identity or secret. OAuth security guidance requires exact redirect matching and protection against open redirects and CSRF.
| Limit | Default and ceiling |
|---|---|
| Attempt, state and custody capsule lifetime | 10 minutes; no extension after exchange starts |
| Callback attempts | 10/session/minute, 60/source-address/minute and 5 failed claims/attempt; exceeding an attempt limit quarantines the attempt rather than bypassing session checks |
| Metadata body | 64 KiB wire content and decoded content, bounded before buffering/parsing |
| Token/registration response | 16 KiB wire content and decoded content, bounded before buffering/parsing |
| Discovery redirects | At most 3, same exact HTTPS origin; validate each resolved address before connection |
| Discovery/token connect and total deadline | 5 seconds and 15 seconds |
| Scope step-up initiation | At most 2/connection/day; a user starts the new consent explicitly |
Count per-attempt callback failures only after authenticating the initiating session and validating signed state. Anonymous/wrong-initiator traffic is source-throttled and cannot consume another user's attempt budget.
Discovery, token and registration HTTP clients request Accept-Encoding: identity, disable automatic decompression and reject every Content-Encoding field, not just bodies that expand beyond a decoded limit. This makes content-byte limits apply before buffering with no decompressor-work loophole. Header sections are limited to 16 KiB and 128 fields. Framing validation and the same limits apply to every redirect response and error body, not just a final success. A provider requiring compressed control responses needs a separately reviewed profile.
For every connection/retry/address fallback, reject private, loopback, link-local, multicast, reserved and cloud-metadata destinations; validate IPv4 and IPv6. Resolve once, validate all candidate addresses and bind the socket to an approved address while preserving the approved hostname for SNI and certificate validation. No second unchecked lookup or redirect can select a different destination. This follows OWASP SSRF prevention guidance.
Custody and lifecycle
Seal token sets, staged material, capsules and registration credentials for every tenant under the existing ADR0083 credential/sealing scope. The connector drafts refine the named sensitive-stream classes; they do not silently change the status or scope of accepted ADRs. Broker replies carrying material remain directly sealed to the authorized caller and never become persistent RPC/idempotency results.
Offer revoke-all before a customer destroys a key, but never delay or pretend to undo the customer's destructive action. If a key veto prevents provider revocation, record that limitation safely and refuse local use. Destroyed/restored custody never authorizes automatic refresh or relay replay. Only reviewed releases may declare stable refresh tokens or reuse grace; discovered servers use the strict rotating-token default.
Project archive suspends use immediately. If not restored within 7 days, start disconnect/revoke cleanup; the tenant may configure a shorter window. Membership and tenant offboarding refuse new use immediately through their authoritative facts. A claim is released only after its Connection is terminal, never because a hold timer expired. Profile erasure starts immediately on disconnect/offboarding; debt is visible until every revision/copy in scope is purged. No numeric cleanup SLA substitutes for completed erasure evidence.
Future person delegations bind exact AgentIds, expire after 24 hours by default and cannot exceed 7 days. Renewal is a new owner-authorized action. Node subjects cannot carry personal delegations. Workspace integration-project binding belongs to workspace policy and governs creation; it does not move existing custody.
Operator authority and retention
Connection upgrades are owner-initiated, including a non-widening release change. Ordinary platform release blocking/app revocation requires a recorded typed reason and approval from distinct platform-operator principals. An emergency operator may block immediately to contain compromise; the durable action requires independent review within 24 hours. Missing review escalates and never automatically unblocks the resource. Operators cannot use this role to read customer credentials.
Retain body-free connection audit facts for 30 days by default. A deployment must explicitly configure a shorter privacy window or a longer justified retention/hold policy before admitting affected tenants; there is no silent indefinite retention. Retention applies to stream replicas, exports and backups through a documented deletion schedule. A legal or contractual hold is an external configured policy, not an exception invented by the proxy. Log-management standards describe policy and operational processes rather than prescribing our chosen day count. NIST SP800-92 is the process reference; this profile supplies the product default.
Keep request/response bodies, query strings, callback URLs with parameters, signature headers and raw provider errors out of logs, traces and audit. Security metrics use bounded reason codes, workload/project references and counts. Alert on replay attempts, denied custody subjects, scope widening, stale authorization, restore epoch mismatch, cleanup debt and exhausted budgets.
Enablement evidence
App registration/approvals, configured DNS/callback certificates, workload identities, network policy, sandbox support and retention/backup settings must be verified in the deployment. Their absence is a concrete enablement blocker; it does not reopen the contract defaults above. No production approval, certification, runtime isolation or vendor registration is claimed by this page.
