Skip to main content
REALIZED.DEV
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.dev is designed to reduce those failure modes through how work is structured, not just through policy.

Book a security review Compliance evidence matrix
Security & Deployment Reference

Runtime pillars

Four pillars, designed to sit outside the model.

The LLM is treated as an untrusted planner. Actions are designed to pass through checks outside the model, and a person approves every merge.

  • Policy checks outside the model

    Rules live outside the LLM, in your repository, and are designed so prompt content does not change them; any change to them is a draft pull request that a person approves.

    • 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

  • Human approval; rollback by reverting the merge

    Agents open draft pull requests. A person approves every merge.

    • Every merge is a person's decision
    • Agents open draft pull requests. A person approves every merge.
    • Rollback: revert the merge to restore the previous state

  • 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
    • Every approval is a merge recorded in Git history

  • Audit Trail in Git

    Every change reaches your main branch as a reviewed pull request, so your Git history shows what changed, when, and who merged it.

    • One pull request per change, with its diff and review
    • Each merge records who approved the change
    • Specs, work items and findings versioned in docs/

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 checks sit outside the LLM. Untrusted content is tagged at ingest, and the workflow is designed so it does not widen its own tool scope.
Credential theft via tool misuse
Access runs through your own repository and Cursor accounts, with credentials designed to stay out of prompts and the repository.
Unbounded blast radius
Write paths are designed to stay within a capability bound. HITL gates pause high-risk classes; a run is designed to stop on an 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
We ship no model. The harness's rules and prompts live in your repository, and every change to them is a reviewed pull request that a person approves.
Missing regulator evidence
Evidence is your Git history of reviewed pull requests. We can walk through how it maps to the EU AI Act and the OWASP Agentic Top 10 (2026). There is no SOC 2 report, ISO certificate or evidence pack to share. Mapped · not certified or audited.

Keys, identity, and where code lives

Your repositories. Your accounts. Your merges.

We don't host a runtime. The work runs in your own repositories and toolchain, using Cursor, and a person on your team approves every merge.

  • Your accounts, your identity

    Access runs through your own repository and Cursor accounts. The harness keeps no credentials in the repository.

  • Where your code lives

    In your own repositories. The harness ships no hosted runtime and no model, so your code goes only where your own Cursor and Git setup sends it.

  • Change control

    Every change, including policy and configuration changes, reaches your main branch as a reviewed pull request that a person approves.

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