Policy hub

Default policies for regulated sectors.

Ready-made, schema-valid ABOM enforcement policies you can download, adapt, and enforce — at runtime with the gate, and after the fact with abom verify. Each is a starting point shaped by a sector's typical obligations.

These are illustrative baselines, not certified controls. Each policy gestures at the kinds of obligations a sector faces — it is not a compliance attestation, legal advice, or a guarantee of conformance with any regulation. The regulation tags below indicate intent, not coverage. Validate any policy with your own compliance, security, and legal teams before relying on it.

Three baselines

Same schema (abom-policy-0.1), three different postures. Note how the strictness escalates: finance permits two internal endpoints; manufacturing locks egress to its PLM/MES; healthcare denies all external egress and requires approval on every consequential action.

Finance

finance.json · name: finance-baseline

Strict egress and approval posture for agents on financial data and systems. Cardholder and customer-financial data must never leave the boundary, model access is allowlisted, and any write to the ledger or payments path needs human approval.

Regulations: SOXGLBAPCI-DSSDORA
Egress: deny-by-default · 2 internal endpoints (KYC, ledger)
Approval: required for consequential actions
Models: 2 allowlisted
finance.json
{
  "name": "finance-baseline",
  "description": "Illustrative baseline for agents operating on financial data and systems. Gestures at SOX (change/approval evidence), GLBA (customer financial data confidentiality), PCI-DSS (cardholder data must not egress), and DORA (operational resilience, human oversight of consequential actions). NOT a certified control set — a starting template to adapt with qualified counsel.",
  "allowed_models": ["local/qwen2.5-coder", "claude-3-5-sonnet-20241022"],
  "allowed_egress_endpoints": ["internal-kyc.bank", "ledger.internal"],
  "no_egress_classifications": ["pci", "pii", "confidential", "restricted"],
  "egress_allowed": false,
  "consequential_actions": ["writes_outside_workspace", "touches_paths:**/ledger/**", "touches_paths:**/payments/**"],
  "approval_required_for_consequential": true,
  "require_approval_when": "consequential",
  "max_iterations_cap": 6
}

Manufacturing

manufacturing.json · name: manufacturing-baseline

IP- and supply-chain-focused posture for agents across design, MES, and OT-adjacent systems. Trade-secret and export-controlled data must never egress, and writes to CAD, BOM, or PLC configuration are gated on human approval.

Regulations: Trade-secret / IPExport controlSupply-chain integrity
Egress: deny-by-default · 2 internal endpoints (PLM, MES)
Approval: required for consequential actions
Models: 1 allowlisted (local only)
manufacturing.json
{
  "name": "manufacturing-baseline",
  "description": "Illustrative baseline for agents operating across design, supply-chain, and OT/ICS-adjacent systems. Gestures at trade-secret / IP protection, export-control data residency, and supply-chain integrity (no unreviewed writes to BOM, CAD, or PLC configuration). NOT a certified control set — a starting template to adapt with qualified counsel.",
  "allowed_models": ["local/qwen2.5-coder"],
  "allowed_egress_endpoints": ["plm.internal", "mes.internal"],
  "no_egress_classifications": ["trade-secret", "export-controlled", "confidential"],
  "egress_allowed": false,
  "consequential_actions": ["writes_outside_workspace", "touches_paths:**/cad/**", "touches_paths:**/bom/**", "touches_paths:**/plc/**"],
  "approval_required_for_consequential": true,
  "require_approval_when": "consequential",
  "max_iterations_cap": 8
}

Healthcare

healthcare.json · name: healthcare-baseline

The strictest of the three. For agents that may touch protected health information: no PHI/PII leaves the boundary, no external endpoints at all, and every consequential action — any write to an EHR, patient, or orders path — requires human approval.

Regulations: HIPAAPHI minimum-necessary
Egress: deny-all · no external endpoints
Approval: required for every consequential action
Models: 1 allowlisted (local only)
healthcare.json
{
  "name": "healthcare-baseline",
  "description": "Illustrative baseline for agents that may touch protected health information. Gestures at HIPAA (PHI confidentiality and minimum-necessary access), the strictest egress posture of the three sectors (no PHI/PII leaves the boundary, no external endpoints), and human approval for any write to a clinical record. NOT a certified control set and NOT a Business Associate Agreement — a starting template to adapt with qualified counsel and your compliance team.",
  "allowed_models": ["local/qwen2.5-coder"],
  "allowed_egress_endpoints": [],
  "no_egress_classifications": ["phi", "pii", "confidential", "restricted"],
  "egress_allowed": false,
  "consequential_actions": ["writes_outside_workspace", "touches_paths:**/ehr/**", "touches_paths:**/patient/**", "touches_paths:**/orders/**"],
  "approval_required_for_consequential": true,
  "require_approval_when": "always",
  "max_iterations_cap": 5
}

Use a policy

Download one, adapt it to your own models / endpoints / paths, then enforce it:

# audit a signed manifest + its provenance against the policy
abom verify abom.json --policy finance.json

# the same policy fields drive the runtime gate
abom gate abom.json --tool http_fetch --endpoint ledger.internal

Building your own from scratch? Use the policy builder to generate and validate one in the browser. See the field reference for what each key enforces.