On 29 September 2026, OpenAI, Anthropic, Google, Meta and Nvidia, reportedly together with xAI, signed a voluntary accord at the White House on safety controls for advanced AI. As MediaNama reported, the companies committed to internal controls that monitor model capabilities in areas such as cybersecurity, biosecurity and chemical threats, an internal oversight team to confirm controls work and issues are fixed, an independent external auditor or evaluator to carry out assessments, and an independent board committee to receive reports.
Our Crest.Digital News Bite explains why independent AI assurance may become a lasting part of enterprise governance. Watch it below, then read on for the practical lifecycle.
The accord has limits worth stating plainly. Reporting indicates there are no penalties for non-compliance, no requirement to publish audit results or auditor names, and no duty to notify government of incidents. It is a statement of intent, not a regulation. The significance lies in the direction it signals for everyone who builds, buys or deploys AI.
Crest.Digital’s Agentic GRC builds configurable AI agents for control testing, evidence collection and remediation tracking, with human-in-the-loop oversight and an audit trail on every action.
Explore Agentic GRCWhat the Accord Says, and What It Does Not
Read closely, the four commitments form a small assurance model: controls inside the company, an internal team that checks them, an outside party that tests them independently, and a board committee that receives the results. That is the familiar three-lines structure applied to frontier AI. What is missing is equally instructive: transparency about what the external reviews found, and consequences if they find problems.
Voluntary frameworks have a way of becoming expectations. Customers write them into contracts, insurers into questionnaires and regulators into guidance. Enterprises that buy or embed these models should expect due diligence questions to evolve from “Do you have AI safety controls?” to “Who tested them, when, and what did they find?” We explored the same pattern in our piece on Microsoft’s AI code of conduct and human control.
Why This Matters Beyond the Labs
Most enterprises are not training frontier models. They deploy AI and agents inside finance, procurement, customer service and security operations, often through vendors. Their exposure is different, but the question a board or auditor will ask is the same: how do you know your AI controls work?
Three gaps usually appear. First, the inventory is incomplete: agents, copilots and embedded vendor AI are adopted faster than they are registered. Second, controls are policy statements rather than tested mechanisms, the gap we discuss in the agentic AI governance gap in audit testing. Third, testing is done by the people who built the system, which is useful engineering but not independent assurance.
Frameworks such as the NIST AI Risk Management Framework give a common vocabulary for AI risk. Independent assurance is what turns that vocabulary into evidence.
A Practical AI Assurance Lifecycle
A workable lifecycle does not need to be heavy. It needs to be complete, so that every AI system or agent can be traced from its existence to the evidence that its controls were tested. The eight stages below do that.
AI and Agent Inventory
List every AI system and agent in use, including embedded vendor AI, with an owner, purpose, data access and the actions each can take. You cannot test what you have not found, as our note on AI inventory before AI governance explains.
Risk Tier
Classify each item by the consequence of failure, not by the technology. Tier decides how much control, testing and oversight it needs.
Controls
Define controls proportionate to each tier: human oversight, access limits, logging, change management and security. Policy alone is not a control, a point we make in why an AI policy is not an internal control.
Independent Testing
Have someone outside the team that runs the system test whether the controls operate. See our view on the independent assurance layer for agent testing.
Findings
Record exceptions with severity, root cause and an accountable owner, rather than letting them live in email and meeting notes.
Remediation
Fix the cause and set a due date, with escalation if it slips. Remediation that is not tracked tends not to happen.
Retest
Test again after the fix. A finding is closed only when the retest passes.
Evidence
Keep the inventory, tier, control, test, finding, fix and retest together as one reproducible trail that a board, auditor or regulator can follow.
The test is traceability. Pick any agent in production and see whether you can show its owner, its tier, the controls that apply, who tested them, what was found, what was fixed and whether the retest passed. If any link is missing, so is the assurance.
Where Agentic AI Helps Without Undermining Independence
There is a natural irony in using AI agents to assure AI, and it deserves care. Agents are well suited to the groundwork: keeping the inventory current, collecting evidence, running scheduled control tests, drafting exception notes and chasing remediation. What they must not do is mark their own homework. Separation of duties still applies, and every agent action should leave a log that a third party can reproduce, as we argue in why AI agents need audit trails.
Crest.Digital’s Voice AI for GRC follows up with control owners for evidence and remediation updates, while the TPRM platform extends the same assurance discipline to AI vendors.
Do not forget the supply side. Many of an enterprise’s AI dependencies sit with vendors and model providers, which makes agentic AI vendor monitoring and independent assurance evidence part of third-party due diligence as well as internal audit.
A Practical Playbook for Boards, CAEs and CISOs
Six Moves to Get Ready for Independent AI Review
- Inventory first: Register every AI system and agent, including embedded vendor AI, with an owner and a list of what it can access and change.
- Tier by consequence: Focus testing on the agents that touch money, customers, regulated data or critical processes.
- Convert policy to controls: For each tier, define a control that can be tested, not only a statement of intent.
- Choose an independent tester: Internal audit, a second line function or an external party, but never the team that operates the system.
- Close the loop: Track findings to remediation and retest, and keep the evidence together.
- Report upward: Give the board or audit committee a regular view of coverage, findings and closure, mirroring the independent committee reporting in the accord.
A final note on proportion. Independent assurance does not mean every chatbot needs a full external audit. It means the depth of testing matches the consequence of failure, and that you can show your reasoning. Our business impact page describes how organizations see results from evidence-led governance.
Crest Perspective
AI governance programmes that treat AI risk as a technology problem will be underprepared for the regulatory environment taking shape in 2025–26. The EU AI Act, NIST AI RMF, and emerging FCA guidance all frame AI risk as a governance and accountability question first — the technology controls follow from having clear answers to who is responsible, what decisions the AI influences, and what the escalation path is when the system behaves unexpectedly.
Frequently Asked Questions
On 29 September 2026, leading AI companies including OpenAI, Anthropic, Google, Meta and Nvidia signed a voluntary accord, reportedly alongside xAI, at the White House. According to press reporting, the commitments cover internal controls that monitor model capabilities around cybersecurity, biosecurity and chemical threats, an internal oversight team that checks controls work and issues are remediated, partnership with an independent external auditor or evaluator, and an independent board committee to receive reports. The accord is voluntary and reportedly has no enforcement mechanism or requirement to publish audit results.
No. Reports describe it as voluntary, with no penalties for non-compliance and no obligation to publish audit results or auditor names, although it was described in public remarks as morally binding and future codification into law was not ruled out. For enterprises the practical significance is not legal force but expectation: if the companies building frontier models accept independent testing of their controls, customers, boards and regulators may start to ask the same of organizations that deploy AI.
Independent AI assurance is testing of AI-related controls by someone who did not design or operate them, with the results documented and any findings tracked to remediation and retest. It differs from a self-assessment or a policy review because the tester has no stake in the outcome. For enterprise AI it typically covers whether systems and agents are inventoried, risk tiered, governed by defined controls, supervised by humans where needed, secured with appropriate access, and supported by evidence that an auditor can reproduce.
Start with an inventory of AI systems and agents, including those embedded in vendor products, then assign each a risk tier so effort goes where consequence is highest. Document the controls that apply to each tier, decide who will test them independently, and make sure evidence is captured as the work happens. Plan for findings: every exception needs an owner, a deadline and a retest. Teams that can walk a reviewer from an agent to its controls, tests, findings and closure evidence are in a far stronger position than those assembling a file after the request arrives.
Agents can do the repetitive groundwork: maintaining the inventory, collecting evidence, running scheduled control tests, drafting exception notes and tracking remediation. Independence is preserved by keeping a separation of duties: the people or agents that operate an AI system should not be the ones that certify its controls, and every agent action should be logged so a third party can reproduce the result. Crest.Digital's Agentic GRC applies this human-in-the-loop model, with accountable people making the judgements and signing off.