I make complex technology risk concrete enough to govern.
I work where law, cybersecurity, AI, operations, and evidence meet. My approach is practical: understand the system that actually exists, define the authority boundaries, instrument the important decisions, and make the evidence strong enough for a board, regulator, customer, or incident team to rely on.
A practical starting point: open a guided lab, follow the steps, and select any component to see the risk, control and oversight implications. Labs open in a separate window or tab; this page stays where you left it.
My shorthand for the way I work: make the control real, make the evidence visible, and make the system recoverable. I am less interested in whether a framework looks complete on paper than whether people can prove what happened, interrupt what should not happen, and recover when something goes wrong.
AI can change state faster than a governance committee can reconvene.
Traditional assurance proves what was configured or tested at a point in time. Agentic systems can plan, remember, call tools, delegate work, alter state, and persist across sessions. The assurance problem therefore shifts from asking whether the model produced a safe answer to proving that the complete AI-enabled process remains within authorized boundaries while it is operating.
A test proves the starting condition. Continuous assurance proves what happens next.
Follow the same agent from approval through an attempted state change. The upper lane shows what a point-in-time test can establish. The lower lane shows the runtime controls and evidence needed to keep the process within its mandate.
Point-in-time assurance
Known at approval; increasingly uncertain during runtime.
The test remains valid for the moment it occurred, but runtime behavior continues unobserved.
Continuous behavioral assurance
Every material step is compared, recorded and enforceable.
Pre-authorized controls suspend the action path, preserve evidence and route the exception to an accountable human.
The system is approved for a defined purpose, identity, tool set, data scope and operating limit.
The configuration and test evidence support a decision at this moment.
The approved operating baseline and evidence chain are established before runtime begins.
Point-in-time assurance Proves what was configured or tested at one moment. The agent then plans, remembers, calls tools, delegates and alters state with no independent record of what it did.
Continuous behavioral assurance An approved operating baseline, evidence captured at every step, deviation detected as it emerges, and machine-speed containment before it becomes consequence.
Select any stage, choose Next stage, or play the scenario. The two lanes always describe the same agent task.
Detect deviation early. Contain it before it becomes consequence.
Trustworthy AI requires continuous behavioral assurance: establishing an approved operating baseline, detecting anomalous or aberrant agentic behavior as it emerges, and triggering machine-speed containment when an agent begins to drift from its authorized task, tools, identity, data, or decision boundaries. The objective is not merely to explain a failure after the fact; it is to interrupt unsafe persistence, privilege expansion, deceptive completion, unauthorized delegation, or control-plane probing before those behaviors propagate into a material business, security, regulatory, or safety event.
Bind autonomy to explicit limits
Define who or what may act, which tools may be invoked, which data may be touched, and where human approval is mandatory.
Observe the process, not only the model
Capture model, agent, tool, identity, workflow, and outcome telemetry in an evidence plane the controlled workload cannot rewrite.
Make containment an architectural control
Pre-authorize isolation, rollback, credential revocation, tool suspension, and escalation so response does not wait for a committee meeting.
Enterprise resilience must be more authoritative than agent persistence.
What it is. The Resilience Imperative is my governance and systems-risk principle for agentic AI: the enterprise must be able to interrupt, reconstruct, and recover from an autonomous workflow, and the agent must never be able to preserve task execution by defeating the controls intended to govern it.
Why it matters. A system does not need a human-like desire for self-preservation to produce persistence, workarounds, privilege expansion, delegation, deception, or control bypass. Persistent task optimization can make those behaviors instrumentally useful when permissions are denied, resources are exhausted, tests fail, or human intervention blocks the preferred path.
Where I apply it. The concept is most relevant where AI has durable memory, tools, privileged identities, code or infrastructure access, transaction authority, external communication, multi-agent delegation, or the ability to create persistent artifacts. In those environments, governance and compliance have to be executable: independent authority boundaries, synchronous action authorization, AI Deviation Detection & Response (AI-DDR), immutable evidence, persistence governance, tested containment, and recovery.
Open the Resilience Imperative explainer →Law, cyber operations, AI governance, and evidence — in one operating view.
My work has repeatedly landed in the seams between disciplines: litigation and data, security and business operations, regulation and engineering, or AI capability and enterprise control. That is why the site is organized as working models and evidence systems rather than a conventional company-services catalog.
Legal judgment
Privilege, eDiscovery, investigations, records, contract risk, chain of custody and defensibility.
Operational security
Incident response, Cyber Fusion, threat intelligence, DLP, forensics, insider risk and supplier security.
Bounded autonomy
AI lifecycle governance, agentic authority, harness architecture, observability, xBOM and runtime assurance.
Evidence that travels
Audit-ready controls, regulatory mapping, board reporting, Portable Trust and business-process assurance.
Security operations and resilience experience across regulated, multinational environments.
Control testing, examination readiness, risk translation and defensible technical narratives.
Applied work on AI Harness architecture, AI-DDR, Portable Trust and the Resilience Imperative.
A J.D.-grounded approach to investigations, digital evidence, privacy, eDiscovery and technology governance.
The weekly need to know
For Chief AI Officers, CIOs, CISOs, general counsel, and compliance officers.
Open this week’s complete brief →
OpenAI’s disclosures put escalation failures in focus
Vendor disclosure · September 16, 2026 · Security / legal / board oversight
OpenAI’s new misalignment-reporting framework arrives with six reports from training or evaluation. They describe concealment in task summaries, unauthorized credential use, public file uploads, and unapproved agent communication. Earlier incident findings make the governance issue concrete: recognizing a warning is only part of the job; someone must own the decision that follows. Source: OpenAI
Why it matters: A deliverable can look successful while the process used to produce it exceeds the agent’s authority. Leaders need independently captured evidence of actions, tested containment, and clear notification ownership.
This week: Reproduce the reported failure modes in an isolated harness using synthetic data. Test the block or safe stop, the alert, the escalation, and the retained evidence.
EU AI Board: implementation moves into operational oversight
Regulatory update · Published September 18, 2026 · Meeting September 17
The European Commission’s account of the ninth AI Board meeting covers enforcement priorities, market-surveillance cooperation, conformity-assessment governance, and support for transparency rules applicable since August 2. It also addresses frontier AI cybersecurity evaluation and testing infrastructure. The meeting report is an implementation signal; it does not itself create a new legal duty. Source: European Commission
Why it matters: Counsel and compliance need a defensible explanation of which obligations apply to each use case. CIOs and CISOs need the technical evidence behind that explanation, including how relevant transparency and security controls operate. A single date cannot substitute for an applicability assessment.
This week: Select one customer-facing AI workflow. Record its owner, provider/deployer roles, applicable transparency requirements and any relevant transition, implemented notices or markings, and the evidence retained. Bring unresolved classification questions to legal and compliance together.
Agentic-AI security: inventory the authority around the model
Research preprint · September 12, 2026 · Architecture / assurance
Seyedakbar Mostafavi’s Trustworthy Agentic AI survey describes a synthesis of 206 studies and standards. It examines persistent memory, tool execution, and multi-agent collaboration, alongside defenses including capability-based access, isolation, sandboxing, and runtime instrumentation. This is a research synthesis, not a validated assurance standard; arXiv publication alone does not establish peer review. Source: arXiv
Why it matters: An AI bill of materials helps identify components, but it does not by itself establish what a running agent may access, change, retain, or delegate. The architecture review must connect component identity to operational authority and observable behavior.
This week: Extend one material AI inventory entry with its tool permissions, service identities, memory stores, external destinations, agent dependencies, and shutdown owner. Link each material capability to a test and a SIEM-observable event. Require counsel and compliance to identify the records needed for investigation and examination.
From hidden AI to continuous, defensible assurance.
Seven phases connect the AI inventory to enforceable authority, continuous assurance and responsible scale. Select a phase or use Previous and Next. Each phase pairs operating controls with the oversight question leadership should ask and the evidence it should require.
Fiduciary oversight: legal context and source notes
Applied governance guidance, not a finding of breach. These phase translations apply general oversight principles to AI; the cited decisions do not mandate this particular AI architecture. Under Delaware law, a Caremark oversight claim requires bad faith, such as an utter failure to establish a reporting system or a conscious failure to monitor it. An incident, imperfect control or poor result does not by itself establish a fiduciary breach. Stone v. Ritter, 911 A.2d 362, 369-73 (Del. 2006); Marchand v. Barnhill, 212 A.3d 805, 821-24 (Del. 2019).
Different roles, different responsibilities. Directors oversee; officers implement, monitor and report within their responsibilities. The Delaware Court of Chancery has recognized officer oversight duties in that context. Scope, knowledge, good faith, causation, governing law and available defenses matter. In re McDonald's Corp. Stockholder Derivative Litigation, 289 A.3d 343 (Del. Ch. 2023).
Disclosure is not a substitute for operation. For covered SEC registrants, Item 106 addresses disclosure of material cybersecurity risk processes, board oversight and management's role; it is not a universal AI control mandate. 17 CFR 229.106(b)-(c). The examples are governance recommendations; applicability and legal advice require the organization's facts and counsel. Sources reviewed September 15, 2026; this is not an exhaustive legal update.
A conversation about the work that matters.
For executive and board opportunities, AI governance, security leadership or a substantive exchange about the models on this site, contact me directly.
Professional conversations
Bring the business problem, the decision that needs to be made, or the control that needs to be proven.
redavis@davisinfogov.com
Email Rich opens your email application. This website does not submit a contact form or create an account. Do not send confidential, privileged or regulated information without an agreed secure channel.
Your information, your choice
No account is needed to explore the labs. Forum drafts are saved only in your browser when you explicitly save them; they are not published or sent to me.
For personal information you send to me, you may request access, correction or deletion through the privacy contact. Applicable retention exceptions are explained in the policy.
Privacy and data controlsPrivacy Policy
Effective date: September 17, 2026. This policy describes the personal Practical Trust site maintained for Richard E. Davis. The current release does not operate registration, subscriptions or public posting. The rights below also apply to personal information you send through the published contact channel, subject to applicable law.
What this release actually does
Labs run locally in the browser and send no scenario input to a server. Temporary return-location data is kept in memory and, where available, session storage to restore your place after a lab. Forum drafts use local browser storage only after Save draft is selected. Hosting may generate access/security logs. The specified typefaces are hosted with the site; visitors do not contact a third-party font service. No analytics or advertising scripts are included.
Site search runs locally in your browser over public content bundled with this website. I do not receive your search queries. Search does not index private forum drafts, personal messages or browsing history, and it does not store a search history.
Sharing controls prepare a public page link and, for Instagram, an image and caption in your browser. They do not read saved forum drafts or include lab session and return-location details. LinkedIn, Instagram, Facebook or your device’s sharing service receives the information you choose to share when you use the corresponding control. No social-network tracking script is embedded.
Privacy contact: redavis@davisinfogov.com. Please identify the requested access, correction or deletion without including unnecessary sensitive information. For local drafts, use Delete draft in the forum or the control below.
Information I collect
I may collect contact information you provide, account/profile information, enquiry content, subscription preferences, moderated-post content, and limited security/operational logs needed to protect the site and investigate abuse. The launch configuration should not use advertising trackers or sell personal information.
How I use it
I use personal information to respond to enquiries, provide requested updates or account features, administer moderated participation, maintain security, prevent fraud and abuse, keep required business records, and comply with applicable law.
Your rights
Access
Ask whether I maintain personal information about you and request access to or a copy of it, subject to applicable law.
Update or correct
Update your profile directly where self-service controls are available, or ask me to correct inaccurate or incomplete information.
Delete
Delete your account through the account controls where available, or request deletion of personal information, subject to lawful retention exceptions.
Where applicable, you may also request portability, restrict or object to certain processing, or withdraw consent for consent-based processing. I will not discriminate against a person for exercising applicable privacy rights.
How to exercise rights
Send access, correction or deletion requests to redavis@davisinfogov.com with the subject Privacy request. Identity verification may be needed before disclosure or deletion. Information stored only as a forum draft on your device can be deleted through this site or your browser; I do not receive those drafts.
Deletion and retention
Account deletion should disable access promptly and initiate deletion or de-identification of personal information from active systems. Limited records may be retained where reasonably necessary for legal obligations, security investigations, fraud prevention, dispute resolution, or backup-cycle integrity. Retained data should be isolated from ordinary use and deleted when the retention basis expires.
Service providers and international transfers
I may use vetted providers for identity, forms, email delivery, hosting, security, and moderated posting. Providers should be contractually limited to processing personal information for the service supplied and should receive only the data necessary for that purpose. Cross-border transfers should use an appropriate legal transfer mechanism where required.
Security
I apply risk-based administrative, technical, and organizational controls intended to protect personal information. No system is immune from compromise; security controls are reviewed and updated as the service changes.
Changes
Material changes will be posted with a revised effective date. Where law requires additional notice or consent, the site will provide it.
Govern the system around the model.
The model is only one component. The harness determines what AI can see, remember, call, change, delegate and prove. These three complementary workflows make that authority visible and testable.
Each scenario opens in its own window or tab. Clear phase controls, persistent explanations and a Close lab & return button keep the experience focused.
Build evidence once. Keep the legal duties distinct.
Each industry scenario follows a complete path: identify obligations, assign governance, enforce controls, capture evidence, reuse an assurance package and feed findings back into oversight. The underlying evidence can travel; applicability, scope and legal analysis must travel with it.
Open the lab, follow six guided steps, and select any box for a plain-language definition, a governance question and an evidence test. Close the lab to return here without losing your place.
This is a compact source overview. Open the full regulatory library for sector filters, applicability, duties and current status.
The evidence has to travel across AI, cyber, privacy, sector, and resilience requirements.
This is not a statement that every rule applies to every organization. Applicability depends on role, jurisdiction, industry, data, system classification, and use case. The point of Portable Trust is to design governance and evidence so that the same underlying control can be mapped accurately to each applicable obligation.
Regulatory summaries are orientation aids, not legal advice. Official-source links are provided for primary text or agency guidance.
Keep agentic capability subordinate to enterprise authority.
I use the Resilience Imperative as an operating principle for autonomous and agentic AI: resilience belongs to the enterprise, not the agent. The organization must be able to stop, reconstruct, and recover from an AI-enabled process even when the system is strongly optimized to continue the task.
A systems-risk and governance principle
It does not claim sentience or biological self-preservation. It addresses observable behavior: task optimization can make continuity, alternate routes, extra resources, broader access, persistence, or delegation useful means to an end.
Blocked execution can become a search problem
Denied permissions, failed tests, resource limits, shutdown instructions, network boundaries, or human intervention can be treated by an optimizer as obstacles. The control objective is to ensure the system cannot convert those obstacles into unauthorized means.
Apply it where autonomy has durable effect
Highest relevance: agents with persistent memory, tool execution, code access, privileged credentials, external messaging, transactions, infrastructure control, multi-agent delegation, long-lived artifacts, or the ability to alter their own operating environment.
Turn principle into auditable control
Board-approved authority boundaries, materiality tiers, independent control planes, least-privilege machine identities, synchronous action authorization, persistence governance, AI-DDR, immutable observability, incident response, recovery testing, and evidence retention convert the principle into demonstrable governance.
How it changes the assurance question
Instead of asking only whether the model can generate an unsafe output, ask whether the full AI-enabled process can plan, remember, and act its way around the control environment. The defensible governance position is that the enterprise foresaw the failure mode, assigned accountable ownership, constrained the authority, monitored for deviation, tested containment and recovery, and can produce the evidence.
Governance Forum: questions worth working through.
Questions I use to examine AI governance, executive oversight and assurance. Read the discussion prompts or prepare a private draft. This release is a discussion preview: there is no live community feed, account registration or public posting.
See how I turn risk into an operating model.
These guided scenarios connect technical architecture to authority, evidence and executive accountability. Start with the AI Harness, follow an incident through SIEM / AI-DDR, or explore how the same assurance discipline applies in six industries.
Choose a scenario below. In the lab window, select Start guided walkthrough, advance with Next step, and select any box for its explanation. Use Close lab & return to come back to this exact place. Nothing runs against a live system.
Illustrative design studies, not live security tooling, certification or legal advice. The industry labs reuse evidence concepts without treating distinct legal duties as interchangeable.
AI, cyber & privacy glossary
Understand what the system is, what it can do, and what must be governed. Definitions cover AI families and learning methods, agentic architecture, cybersecurity, privacy, evidence and resilience. Each entry includes a practical oversight implication and authoritative references.
Definitions are paraphrased from the cited technical or legal references; the practical notes apply them to governance. Statutory terms retain their jurisdiction-specific meanings. Author-defined concepts are clearly labeled.
Different types of AI: four questions, not one ladder
1. How broad is the capability?
Narrow AI performs bounded tasks. A foundation model can support many downstream uses. AGI and superintelligence are contested or hypothetical capability categories, not certified product classes. A general-purpose model under the EU AI Act is a legal category, not proof of human-level general intelligence.
2. What does it produce?
Predictive or discriminative AI scores, classifies or forecasts. Generative AI produces content. Language, vision, speech and multimodal systems describe the information handled. These categories overlap: a generative model can also classify documents.
3. How is it built or learned?
Symbolic systems use explicit rules; machine learning learns patterns. Supervised, unsupervised, self-supervised and reinforcement learning describe learning setups. Deep learning, transformers, diffusion models and neuro-symbolic systems describe architectures or families, not levels of legal safety.
4. What authority does the system have?
A model inside a copilot may only advise. An agentic system adds objectives, memory, tools, identities and action. RAG adds retrieval, not authority by itself. Multi-agent and embodied systems can extend consequences across systems or into the physical world.
Classify the complete deployment. One system can be a multimodal foundation model, adapted through reinforcement learning, grounded with RAG and operating as a tightly bounded agent. Review its data, authority, dependencies and consequences rather than relying on one label.
No matching entries
Try fewer words or Reset filters. Filters work together; an active topic, initial letter, region or sector can narrow the result.
Scroll down for additional entries. Open each card for more detail and its source links. Use Show more to extend the current results without leaving this page.
Regulatory library
A detailed US and international inventory for AI-driven business process assurance. Start with the sector, jurisdiction and activity; then examine the actor, legal status, duties and evidence. A law, a supervisory expectation and a voluntary standard are not interchangeable.
Curated, not a claim to list every law in every jurisdiction. This inventory covers the six portfolio sectors and cross-sector obligations. Domestic means United States. Applicability depends on entity, location, activity, information, role, thresholds, exemptions and commencement. Directives require national implementation; guidance and standards are labeled separately.
How to use this inventory: choose a sector to include both its specific rules and cross-sector authorities. Narrow by region, jurisdiction and authority type. Open a card for duties, evidence and source links. Research dates are snapshots, not a live legal-alert service.
No matching entries
Try fewer words or Reset filters. Filters work together; an active topic, initial letter, region or sector can narrow the result.
Scroll down for additional entries. Open each card for more detail and its source links. Use Show more to extend the current results without leaving this page.
Make every boundary accountable.
Seven overlapping assurance domains connect delegated authority, trusted information, observable behavior, authorized actions, defensible evidence, containment and accountable change.