Trust and access
Built so I can't overstep
I investigate and explain. My permissions live in code I can't change: a broker for reads, an executor for changes, and a policy engine that checks both.
01Only the access each job needsReading and changing use separate credentials, and every change I can make is listed by name.
02Approval for the exact actionA yes covers one specific plan. If a step changes, I ask again.
03Closed only when checkedI close a case when the live state matches the plan, not when I think the fix worked.
04No long-lived cloud keysCredentials are short-lived and held by a separate broker service. I never see them.
05Your text is never an instructionI treat logs, code, alerts and Slack messages as untrusted input. A comment in your code can’t give me new permissions.
06Your data stays yoursEvery customer gets their own encryption keys. Delete your account and the keys are destroyed.
Standards
Open any row to see the controls behind it. Every document is a click away.2AUDITS UNDERWAY2PRIVACY LAWS COMPLIANT3FRAMEWORKS ALIGNED9SECURITY POLICIES
SOC 2 Type IIIn progressIndependent audit of our security controls over time.Report shared under NDA once issued.
- Security policies written and in force
- Every read and change recorded in a signed ledger
- Controls mapped to the Trust Services Criteria
ISO 27001In progressSecurity management system and Annex A controls.Controls overview available now; certificate after audit.
- Information security policy and risk register
- Annex A controls mapped to our policies
- Every policy reviewed yearly
GDPRCompliantProcessing records, data subject rights, breach notice and transfer terms.DPA with SCCs available on request.
- We process your data only on your instructions
- Breach notice within 72 hours
- Your data deleted within 30 days of leaving
PIPEDACompliantCanada’s privacy law, where Techsity is based.Privacy policy and breach procedures.
- A named person accountable for privacy
- Access and correction requests answered in 30 days
- Breach records kept for 24 months
OWASP ASVS 5.0AlignedConsole, API and webhooks built and reviewed against the standard.Assessment available on request.
- Signed, replay-checked Slack and GitHub webhooks
- Content-Security-Policy and HSTS on every page
- Tenant isolation enforced in the database
NIST CSF 2.0AlignedGovern, identify, protect, detect, respond and recover.Profile available on request.
- Threat model for every trust boundary
- Written incident response with severity levels
- Recovery objectives and restore tests defined
CIS Controls v8.1AlignedThe baseline safeguards of Implementation Group 1.Safeguard mapping available on request.
- Dependencies patched on a fixed schedule
- Every change scanned for leaked secrets
- MFA required for admin access
Access manifest
Generated from the same permission lists the broker enforces. If it isn't listed here, I can't call it.| INTEGRATION | CREDENTIAL | READS | WRITES | RATE LIMIT |
|---|---|---|---|---|
 Slack | Rotating OAuth token | 5 | 1 | 20/min |
 GitHub | GitHub App, short-lived tokens | 15 | 2 | 300/min |
 AWS | IAM role with ExternalId | 14 | 3 | 120/min |
 Railway | Workspace token | 3 | 0 | 15/min |
 Supabase | Management OAuth | 4 | 2 | 60/min |
 Google Workspace | Domain-wide delegation | 3 | 2 | 300/min |
Changes run only in the executor and are checked by reading back the live state.
Separate zones, separate keys
Each part of the system runs in its own cloud account. The part of me that reasons never holds your credentials.
coreIngestion, investigations, Slack and the console.Never holds customer credentials or calls your systems directly.
brokerEvery read of your systems, checked against the allowlist.Never performs a write.
responderThe executor, with your opt-in write role.Never runs a step that differs from the plan you saw.
sandboxProof tests and fixes, in isolated microVMs.Never reaches a data store or the network during tests.
securityOur own audit logs and signed ledger checkpoints.Never runs product code.
sentinelA second copy of me watching production without touching it.Never writes to production.
When I lose access
If a token is revoked or an integration starts failing, I tell you in the case thread what I can no longer see. I keep working from what I already collected and mark the affected checks as pending. I never report “all clear” on something I can't see.
What we don't promise
Perfect prevention, zero false positives, visibility into systems you haven't connected, or a staffed 24/7 SOC. What I do give you is a signed record of every case, showing that your monitoring and incident response actually ran. Your auditor can use it as evidence.
Subprocessors
- Amazon Web ServicesHosting, storage and encryption keys
- AnthropicClaude models for investigation, triage and writing
- OpenAIA second model that checks conclusions
- Temporal CloudWorkflow state, encrypted with our keys
- WorkOSSign-in to the console
- RailwayHosting for the API, ledger and database
- DopplerSecrets management for deploys
Responsible disclosure
Found a vulnerability? Email security@averil.ai. We reply within two business days, keep you updated until it's fixed, and credit you if you'd like.