Security, Trust & AI Governance
Hobasa does not use client data to train models unless the client explicitly opts in through a written contractual agreement.
By default, client data is used only to deliver the contracted service, remains logically isolated per client, and is never used to improve models serving other customers. Our program is designed around the reality that Hobasa runs both conventional software and autonomous AI agents on behalf of CPA firms and their end clients, so both must be governed to the same standard.
This page is maintained by Hobasa to answer common security, privacy, and AI governance questions about our platform. It is not an independent certification or audit attestation.
Nine principles that shape how we build and operate.
Each principle below is a working control set our engineering and security teams operate against, not a marketing aspiration.
No user, service, model, or agent is trusted by default. Every connection is authenticated and every action is authorized using least-privilege access.
Client data is segregated by tenant, encrypted in transit with TLS 1.3 and at rest with AES-256. We collect and retain only what's required to deliver the service.
PII, PHI, and sensitive data are minimized, redacted, tokenized, or de-identified before being processed by models or agents. Residency and deletion are enforced technically and contractually.
We defend against prompt injection, jailbreaks, model poisoning, extraction, hallucinations, adversarial inputs, and supply-chain risks. Models and artifacts are tracked via an AI Bill of Materials, signed, scanned, and monitored.
Agents are governed differently than chatbots because they take actions. Every agent has a unique identity, short-lived credentials, governed tool access, parameter-level authorization, and immutable logs.
High-impact or irreversible actions always require authenticated human approval. Every autonomous decision is traceable to a goal, policy decision, tool action, and a named accountable human.
Every significant agent action is logged and reconstructable: goal, input, tools used, policy checks, human approval, and outcome. "The AI did it" is not an acceptable answer in our governance model.
Threat modeling, red-teaming, runtime monitoring, drift detection, anomaly alerts, and tiered containment keep systems safe after deployment, not just at launch.
Mapped to NIST AI RMF, ISO/IEC 42001, ISO/IEC 27001, the EU AI Act, and GDPR — plus HIPAA, PCI DSS, and SOX where applicable to the workload.
What we operate, every day.
Continuously operated by Hobasa's security and platform teams and reviewed during our annual audit cycle.
Identity & access
- SSO via SAML 2.0 and OIDC for all paid plans
- Enforced multi-factor authentication for staff
- Role-based access with least-privilege defaults
- Just-in-time access elevation with full audit trail
Data protection
- AES-256 at rest, TLS 1.3 in transit
- Tenant-scoped data isolation per client
- Customer-managed key support on Enterprise
- Field-level redaction and tokenization for sensitive PII/PHI
Infrastructure
- Hardened cloud environments in US regions
- WAF and DDoS protection at the edge
- Continuous patching and automated vulnerability scanning
- Third-party penetration testing on a recurring cadence
Application security
- Mandatory peer review and CI security gates
- SAST, secret scanning, and SBOM on every build
- Annual secure-coding training for engineering
- Coordinated vulnerability disclosure program
Monitoring & response
- 24×7 alerting on anomalous access and data movement
- Centralized, immutable audit logging
- On-call rotation with documented runbooks
- Tabletop exercises run twice per year
Operational resilience
- Automated backups with point-in-time recovery
- Documented RTO of 4 hours and RPO of 1 hour
- Multi-AZ failover for production services
- Annual disaster-recovery testing
How we govern models and autonomous agents.
Because Hobasa runs both GenAI features and agents that can take actions in connected systems, we operate a distinct control set for AI on top of our platform controls.
Every model, dataset, prompt template, and artifact is tracked in an AI Bill of Materials — signed, versioned, scanned for known risks, and monitored in production.
Input and output filters, prompt-injection detection, jailbreak heuristics, PII/PHI redaction, and grounding checks reduce hallucination and unsafe output.
Each agent runs with a unique workload identity, short-lived credentials, and a narrowly scoped tool catalog. Tool calls carry parameter-level authorization checks.
High-impact, financial, or irreversible actions (payments, filings, external notifications, mass data changes) require authenticated human approval before execution.
For every agent action we retain the originating goal, policy evaluation, retrieved context, tool calls, approver, and outcome — reconstructable on demand for audit.
Ongoing red-teaming, adversarial evals, drift detection, and anomaly alerts feed a tiered containment playbook that can throttle, sandbox, or disable an agent quickly.
Clear answers on the data questions firms ask first.
Only financial, workforce, and account metadata your firm connects or uploads — the minimum needed to deliver the requested analysis.
Authorized users in your firm. Hobasa staff access is least-privilege, time-bound, and recorded in the audit log. Agents access only what their scoped tools allow.
Retained for the life of the agreement. On termination, customer data is deleted from production within 30 days and from backups within 90 days, subject to legal retention.
Primary processing in US cloud regions with multi-AZ redundancy. Regional residency available on Enterprise; sub-processors maintained on a published list.
By default, client data is used only to deliver the contracted service. It is never used to train shared models. Training on client data requires an explicit written opt-in.
Tenant isolation, AES-256 at rest, TLS 1.3 in transit, field-level redaction for sensitive PII/PHI, and customer-managed keys on Enterprise.
Frameworks we map to.
Our program is designed around recognized frameworks so your audit, IT, procurement, and risk teams can evaluate us with standards they already use.
NIST AI RMF
AlignedGovern, Map, Measure, Manage functions applied to models and agents.
ISO/IEC 42001
AlignedAI management system principles inform our AI governance program.
ISO/IEC 27001
AlignedInformation security management controls across the platform.
EU AI Act
AlignedRisk-tiered controls, transparency, and human oversight requirements.
GDPR & US state privacy
AlignedDPA available on request; access and deletion honored within 30 days.
HIPAA / PCI DSS / SOX
Where applicableSector controls applied to workloads that fall in scope; BAA available.
SOC 2 Type II
In observation periodTargeted report on our published roadmap.
AICPA confidentiality standards
AlignedBuilt around the workflow needs of CPA firms.
Status descriptions reflect Hobasa's own assessment of program maturity and are not a substitute for a third-party certification. Reports and letters of attestation, where available, are shared under NDA.
A documented playbook, not a draft.
- Detect
Monitoring, alerting, and reports from customers or researchers feed the on-call queue.
- Triage
Severity assigned within 15 minutes; incident commander assembled for Sev-1 and Sev-2.
- Contain
Affected systems isolated; tokens rotated; agents throttled or sandboxed; forensic snapshots captured.
- Notify
Affected customers contacted without undue delay, with a written account once facts are confirmed.
- Learn
Blameless postmortem within 5 business days; corrective actions tracked to completion.
FAQ