Every framework asks
for the same things.

Compliance is fracturing. The requirements converge.

A company operating across the US and EU now faces 24 state privacy laws, GDPR, the EU AI Act, and potentially HIPAA, SOC 2, and PCI DSS — each with its own definitions, deadlines, penalties, and enforcement agencies. The compliance surface is expanding faster than any legal team can track. As of mid-2026, new state laws are taking effect every few months.

But beneath the jurisdictional differences, every framework demands the same core properties. They use different words — "appropriate technical measures" (GDPR), "reasonable security practices" (CCPA), "administrative, physical, and technical safeguards" (HIPAA) — but they all require the same things from the infrastructure.

Ten requirements. Every framework. The question for any company deploying AI agents is not "which requirements apply to us" — all of them do. The question is: does the infrastructure satisfy them structurally, or does someone have to remember to comply?

Every requirement maps to a Safebox property.

Safebox wasn't designed for compliance. It was designed for security — to make compromise structurally evident or impossible. But security and compliance converge: the properties that make a system secure are the same properties that make it compliant. The difference is that Safebox enforces them at the infrastructure layer, which means they hold regardless of what the application, the operator, or the AI agent does.

What Safebox gives you for each regulation.

GDPR (EU — €20M or 4% of global turnover)

GDPR applies to any system processing EU residents' personal data, regardless of where the system is hosted. In 2026, cumulative fines have reached €5.88 billion since the regulation took effect. AI systems face dual compliance obligations as the EU AI Act layers on top of GDPR.

Safebox addresses the GDPR requirements that are hardest for AI deployments: Article 25 (data protection by design — the capability model and controlled egress), Article 30 (records of processing — immutable action streams), Article 32 (security of processing — TPM attestation, ZFS encryption, per-org isolation), Article 35 (DPIA — the attestation claim and workflow documentation serve as the technical input to a DPIA), and Article 22 (automated decision-making — M-of-N governance ensures human oversight on consequential decisions).

EU AI Act (€35M or 7% of global turnover)

The EU AI Act is the world's first comprehensive AI regulation. Article 50 transparency obligations are live as of August 2, 2026. High-risk system obligations were deferred to December 2027 (Annex III) and August 2028 (Annex I) by the Digital Omnibus, but transparency requirements are already enforceable. 78% of organizations had not taken meaningful steps toward compliance as of April 2026.

Safebox maps directly to the AI Act's technical requirements: Article 9 (risk management — governance engine with per-action evaluation), Article 10 (data governance — Streams-based provenance), Article 11-12 (technical documentation and logging — attestation claims and immutable action streams), Article 14 (human oversight — M-of-N governance), and Article 50 (transparency — OCP claims with model version, prompt provenance, and signed execution records).

HIPAA (US healthcare — $2.15M per violation tier)

92.7% of healthcare organizations experienced an AI-agent security incident in the past year. HIPAA requires Business Associate Agreements (BAAs), minimum necessary access, audit controls, and breach notification within 60 days. A proposed 2025 update to the Security Rule would require annual verification of vendor compliance.

Safebox's architecture is a natural fit for HIPAA: minimum necessary access is the capability model (agents access only what they're explicitly granted), audit controls are immutable action streams (§164.312(b)), encryption is TPM-sealed ZFS (§164.312(a)(2)(iv)), access controls are per-org HKDF isolation (§164.312(a)(1)), and breach detection is attestation failure (the PCR measurements change if the image is tampered). A healthcare organization deploying AI agents inside Safeboxes can demonstrate to an auditor that the technical safeguards are structural properties of the infrastructure, not policies that depend on correct configuration.

24 US State Privacy Laws

Alabama, California, Colorado, Connecticut, Delaware, Florida, Indiana, Iowa, Kentucky, Louisiana, Maryland, Minnesota, Montana, Nebraska, New Hampshire, New Jersey, Oklahoma, Oregon, Rhode Island, Tennessee, Texas, Utah, Vermont, and Virginia each have comprehensive consumer privacy legislation. The common requirements: right to access, correct, and delete personal data; opt-out of targeted advertising and data sales; data protection assessments for high-risk processing.

The Streams data model gives you the access, correction, and deletion infrastructure per-user and per-stream, with explicit provenance. The action streams give you the processing records. The capability model gives you the data protection assessment input — you can enumerate, for any given workflow, exactly which data it accesses, which external services it calls, and which governance approvals it requires, because the declarative workflow is statically analyzable.

SOC 2

SOC 2 Trust Services Criteria require security, availability, processing integrity, confidentiality, and privacy controls with documented evidence. Safebox's immutable action streams are the evidence trail. The attestation claims are the system description. The governance engine is the documented control. The TPM attestation is the access control evidence. A SOC 2 auditor reviewing a Safebox deployment sees the controls as architecture, not as screenshots of a dashboard someone checked last quarter.

Structural compliance vs. operational compliance.

Traditional compliance is operational: someone configures encryption, someone writes an access policy, someone sets up logging, someone remembers to review the audit trail. Every "someone" is a point of failure. The 88% AI-agent incident rate in 2026 is what happens when compliance depends on correct human behavior across every deployment, every configuration, every update, every team.

Safebox compliance is structural: the encryption is in the filesystem (ZFS + TPM), the access control is in the capability model (no capability = no action), the audit trail is in the data structure (append-only streams), the human oversight is in the governance engine (M-of-N required, not optional). These properties hold regardless of whether the operator remembers to configure them, because they're not configuration — they're architecture.

The structural approach doesn't replace operational compliance — you still need policies, training, and a DPO. What it does is make the technical controls the part a regulator or auditor can verify independently, without trusting anyone's word. The attestation claim is signed by the TPM. The action streams are cryptographically sealed. The capability model is sha256-verified. An auditor doesn't need to trust the operator — they verify the math.

What's coming, and when the fines start.

The EU AI Act's Article 50 transparency obligations are already live. A company deploying AI agents that interact with EU residents needs to disclose AI use, label AI-generated content, and maintain records. The high-risk obligations — the full conformity assessment, risk management, and documentation requirements — land in December 2027. The infrastructure to support those obligations takes months to build. The 100+ companies that signed the "months, not years" letter are telling you the timeline themselves.

Compliance is an infrastructure decision.

You can hire consultants to write policies, train staff on procedures, and fill out checklists. You'll pass audits — until the next breach, the next misconfiguration, the next employee who forgets to follow the policy. 88% of organizations had an AI-agent security incident in the past year. The policies existed. The policies weren't enough.

Or you can choose infrastructure that satisfies the technical requirements structurally. The encryption is in the filesystem. The audit trail is in the data structure. The access control is in the capability model. The human oversight is in the governance engine. When the auditor asks "how do you ensure encryption at rest?" the answer is "the dataset won't open without the TPM-sealed key" — not "we have a policy that says to enable encryption."

Safebox doesn't replace your compliance program. It makes the technical controls the part you don't have to worry about. The regulatory landscape is fracturing — 24 states, GDPR, the AI Act, HIPAA, SOC 2 — but the technical requirements converge. One substrate. Every framework. The properties that make the system secure are the same properties that make it compliant. Build on the right substrate and the compliance follows from the architecture.