Skip to main content
REALIZED.TEAM
Book Workshop

Security

Governance as architecture.
Not guardrail prompts.

The 2026 threat surface for autonomous agents is not misalignment — it is prompt injection, credential theft, and unbounded blast radius. Realized.Team is engineered so those failure modes are structurally impossible, not policy-forbidden.

Book a security review Compliance evidence matrix

Runtime pillars

Four pillars, enforced beneath the model.

The LLM is treated as an untrusted planner. Every action is mediated by deterministic components that the model cannot override — by architecture, not by prompt.

  • Deterministic Policy Enforcer

    Hard constraints live outside the LLM as compiled rules. Prompt injection cannot rewrite them, because the model never touches the enforcement code path.

    • Compiled from your OWASP ASI 2026 + internal control catalogue
    • Per-action authorization checked before any tool call executes
    • Denials returned to the model as structured feedback, not silent failures

  • Circuit Breaker + Stop Button

    A single Stop Button halts every running workflow in under 60 seconds — the EU AI Act Art. 14 emergency-shutdown mandate, engineered as a substrate primitive.

    • Global, per-workflow, and per-agent kill switches
    • Auto-trip on anomaly signals (error rate, cost, blast-radius heuristics)
    • Rollback to last known-good state at the phase boundary

  • HITL Gate

    High-risk action classes auto-pause for one-click human approval. Uncertainty telemetry surfaces in the UI to counter automation bias before an action commits.

    • Configurable per action class, per environment, per role
    • Uncertainty and provenance scores rendered inline
    • Signed approvals join the audit fabric as first-class evidence

  • Immutable Audit Fabric

    Reasoning chains, tool calls, credentials used, and state diffs are cryptographically signed and append-only. FINMA-ready evidence on day one, exportable on demand.

    • Signed hash chain per workflow instance
    • Regulator-grade export in SIF / OSCAL formats
    • Retention aligned with Swiss FADP Art. 21 and EU AI Act Art. 12

Threat model

The 2026 agentic threat surface, mapped to controls.

Every threat class in the OWASP Agentic Security Initiative 2026 landscape maps to a runtime control — with the evidence exported to the compliance matrix.

Threat class
Runtime mitigation
Prompt injection / instruction hijack
Policy enforcer sits outside the LLM. Untrusted content is trust-tagged at ingest and cannot escalate its own tool scope.
Credential theft via tool misuse
Short-lived, scoped credentials issued per-action by the runtime — never held by the model. Vault integration with per-workflow key isolation.
Unbounded blast radius
Every write path is capability-bounded. HITL gates trip on high-risk classes; circuit breaker kills the workflow on anomaly signal.
Cascading hallucination in long plans
Each Observe → Decide → Execute step closes with a grounded verification pass. Uncertainty telemetry feeds HITL routing.
Model / prompt supply-chain compromise
Model weights, prompts, and connectors are pinned and hash-verified. Change-control signed against the audit fabric.
Missing regulator evidence
Every action produces exportable evidence pre-mapped to EU AI Act Art. 12, FINMA 08/2024, ISO 42001, and SOC 2.

Keys, identity, and residency

Your keys. Your identity plane. Your region.

Deployment topology never affects the governance contract — but it does determine where your keys and data live.

  • BYO KMS + identity

    Runtime encryption keys live in your KMS (AWS KMS, Azure Key Vault, GCP KMS, or HashiCorp Vault). Identity via OIDC / SAML / SCIM.

  • Data residency

    EU, US, CH, and APAC regions. Air-gapped on-prem topology for regulated workloads with no outbound egress from the runtime plane.

  • Change control

    Every policy change, connector scope change, and model swap ships through signed change requests recorded in the audit fabric.

Next step

Walk through the security architecture with our team.

90 minutes with your security and compliance stakeholders. You leave with a scoped control map against OWASP ASI 2026, the EU AI Act, and your internal catalogue.

Book security review Compliance evidence matrix