For Regulated AI, the Rules Are Turning Into Operating Checklists
This week, data-control principles, disconnected deployment playbooks, a hardened machine-learning platform and a new critical-infrastructure code all tell builders exactly what to implement.
- Fivetran proposes five policy principles, including data control, portability and governed data as the basis for AI accountability. m:…7076
- Air-gapped AI means no internet connectivity, no outbound traffic and no runtime dependence on an external provider. m:…2951
- H2O-3 Secure is built to FedRAMP High standards and supports SOC 2, ISO 27001 and ISO 42001, the company says. Self-reported m:…4871
- Singapore's CCoP 2026 requires board accountability, visibility into connected systems and continuous monitoring. m:…5900
Five principles for data and AI
Fivetran has proposed five public policy principles for data and AI, starting with the idea that people should know how their data is used and be able to access or delete it. m:…7076
The remaining principles, as proposed by Fivetran, are that data should stay in the control of those who create and rely on it, that data should be portable rather than trapped in closed formats, that reliable and well-governed data is the foundation for AI accountability, and that privacy rules should protect people without creating prohibitive costs for smaller companies. m:…7076
The urgency comes from what agents now touch: Fivetran notes that modern AI agents can access enterprise data to assist in decision-making, that personal agents are gaining access to sensitive data such as health and financial metrics, and that airline agents can already rebook flights and issue boarding passes without human intervention. m:…7076
Principles alone do not deploy software, and the rest of this week's material shows what compliance looks like once it reaches engineering teams. m:…7076
Disconnected by design
Domino Data Lab describes air-gapped AI as any workload running entirely inside a network boundary, with no internet connectivity, no outbound traffic and no runtime dependence on an external provider. m:…2951
That definition forces a different workflow: deployments must shift from pulling dependencies on demand to pushing everything in advance, with all dependencies staged internally, and Python package mirrors must cover transitive dependencies and GPU-specific builds rather than just the main package index. m:…2951
The cost of isolation is real. Domino cites a Gartner estimate that the sovereign infrastructure premium runs 15 to 20 percent because environments must be duplicated, and notes that sovereignty and isolation requirements can rule out centralised deployment patterns. m:…2951
The payoff, in Domino's account, is speed inside the boundary: the United States Navy's Project AMMO uses the Domino platform to shorten model update cycles for automatic target recognition from months to days. m:…2951
Open-weight models such as Llama, Gemma, Qwen and Mistral are practical for these environments because they can be installed locally, Domino notes, and it prefers safetensors over pickle-based checkpoints for model imports to prevent arbitrary code execution during loading. m:…2951
Supply-chain tooling is catching up: Domino cites an expectation that 85 percent of large enterprise engineering teams will deploy software supply chain security tools by 2028, up from 60 percent in 2025. m:…2951
A hardened platform and a new code
H2O.ai, the company says, has launched H2O-3 Secure, a commercially licensed, production-grade edition of its open-source machine learning platform, built to FedRAMP High engineering standards and designed for financial services, insurance, telecommunications, healthcare and the public sector. m:…4871
The platform, the company says, supports compliance frameworks including SOC 2, ISO 27001 and ISO 42001, includes automated model documentation and built-in drift monitoring, and keeps its open-source edition free under the Apache 2.0 licence. m:…4871
H2O.ai also says its model artifacts score in a few hundred microseconds per prediction, and that the company is trusted by more than 20,000 organisations, including half of the Fortune 500. m:…4871
In Singapore, the updated CCoP 2026 code for critical infrastructure shifts the security focus beyond the designated critical environment itself, emphasising board and senior-management accountability and visibility into systems connected to critical infrastructure. m:…5900
The code requires continuous monitoring and behavioural baselining, distinguishes between enterprise systems, interconnected systems and critical systems, and requires monitoring of data flows between operational technology and enterprise systems for anomalies. m:…5900
Vectra AI, the company says, supports 31 of the 53 clauses in the code, with its strongest alignment in visibility, monitoring, detection, threat hunting, incident response and operational technology monitoring. m:…5900
What is still unresolved
- Fivetran's principles are proposals, not enacted law, and no adoption timeline is given.
- The sovereign infrastructure premium is a Gartner estimate cited by Domino, not a measured cost from a named deployment.
- How many organisations can meet the full CCoP 2026 monitoring requirements with existing tooling is not stated.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.