Architecture overview
Why local execution is safer than the cloud
When a SaaS vendor loses a database, millions of customers' data can be exposed in a single incident. ABN runs as a node on your side of the firewall — your server, your VPC, your machine. There is no central ABN database that can leak. There is no shared infrastructure that can compromise more customers than the one affected by its own incident. This is not a feature; it is a design principle.
You own the execution
The entire ABN stack — Observer, Process Graph, Autonomous Engine, Trust Layer — installs locally. No "SaaS control plane" that can see your data.
No subprocessor sees data
Nango proxies OAuth — not data. The LLM gateway is called only with abstract structures. No third-party subprocessors see customer values.
Zero-Trust
Every agent, every action, validated
Signed blueprints
Every agent blueprint is cryptographically signed. ABN verifies the signature before the agent runs and refuses any agent whose signature is missing or invalid — a blueprint cannot be altered after approval without detection.
Policy-driven execution
A per-tenant policy decides which systems, fields and action classes an agent may touch. Write-guard, rate-limiter and culture rules run on every output.
Stateless design
No data to steal, no history to break
The engines are deliberately without internal memory. Each run reads the state from the database, runs in RAM, writes the result back, and terminates. An attacker who compromises the process at a given moment gets no history — only that run's data packet, which in turn has passed PII Guardian. There is no long-term memory to exfiltrate.
PII Guardian
Automatic PII detection and blocking
Data Minimizer first
Per (system, resource) there is an explicit whitelist of fields. Everything outside the whitelist is discarded before anything reaches memory. Default deny.
SHA-256 pseudonymisation
PII Guardian scans string values and replaces identifying information with stable SHA-256 tokens. Same value → same token, but never the identity.
Result Attestation Layer
Cryptographic proof of data integrity
For every agent output, RAL creates a signed attestation document that binds the finding to the exact source records used. If the source changes during the run, delivery is aborted and the finding is marked stale. You can always ask ABN "which data was used and why?" and get a verifiable answer.
HMAC-signed reports
Every attestation is written to the customer's own `abn_attestations` table. You don't have to trust ABN — you can SQL the proof out yourself.
Transparency tables
Six customer-owned tables log every activity, LLM call, access, attestation, proposal and permission change — for audit whenever you want.
LLM data policy
LLM policy — you choose the level
Choose the level of LLM access that suits your organisation. The setting is per-customer and can be changed at any time in the dashboard — every LLM call is logged to your ownabn_llm_callstable so you can always audit exactly what was sent.
That commitment is not just architectural — it is measured. Before any prompt leaves the node, it is checked to confirm no raw customer value is present, and that payload-clean result is recorded for every single call. The unexpected fails closed: if anything that should not be there is detected, the call is aborted rather than sent.
Default
Redacted
PII is removed automatically. Structural data is sent encrypted to Claude. Never raw values.
Strict
No Data
Only event type names and statistics reach the LLM. No field values at all — neither raw nor anonymised.
Local execution
Local Only
The LLM runs on your own server with Mistral. Not a byte leaves your infrastructure. Suited to banks, healthcare and the public sector.
Fail-closed: if the local model is unavailable the call is aborted — ABN never falls back to a cloud LLM in this mode.
Explicit opt-in
Full
All abstract data reaches the LLM. Requires explicit per-customer approval — never the default.
The four modes are handbook § 3.4. No customer is moved between modes without us talking first — every upgrade or downgrade is logged in the legal centre.
Compliance
GDPR, NIS2, EU AI Act
ABN is built per GDPR Article 25 (privacy by design), ships with a ready Article 28 data processing agreement template and meets the documentation requirements of the EU AI Act Article 11 / Annex IV. Every agent is risk-classified and ships with its own Article 11 / Annex IV technical declaration, generated automatically when the agent is created — compliance is enforced and proven at creation, not assembled after the fact for an audit. NIS2-relevant controls — incident response, security measures, reporting chains — are documented in the legal centre.
What we don't have yet
SOC 2 — roadmap, not done
ABN does not yet have an independent SOC 2 Type 1 attestation.
We are honest about this. The internal security checklist (published in the legal centre) follows the ISO 27001 and SOC 2 Type 1 structure, but a formal audit is planned — not completed. If the contract requires SOC 2, it is a parameter we discuss early and build a timeline around together with you.
Security questions?
We answer concretely
Security questions, penetration tests and technical due-diligence reviews are answered by staff with a technical mandate — not a sales script.