On September 30, six tech executives signed a document at the White House: Sundar Pichai, Dario Amodei, Mark Zuckerberg, Greg Brockman, Elon Musk, and Jensen Huang. The President signed it too.
Most of the coverage filed it under "voluntary." I read the text, and that word stopped doing much work about halfway down the page. Each company commits to internal controls on its models, an internal team that checks those controls actually work, an independent external auditor, and a committee of its board that receives the findings and oversees the fixes.
Pledges don't come with an outside auditor and a board committee. This is an audit regime with a signing ceremony attached.
Once a board has the report, it can't unread it
David Sacks made the point on the All-In podcast two days later. A board committee that sits on an external auditor's findings is exposed: directors carry fiduciary duty, their D&O coverage depends on honoring it, and the FTC and SEC stand behind any public commitment a company makes. Voluntary to sign. Mandatory to act on.
I've spent thirty years building and selling technology into large companies, at Logitech, RAPT, Afero, and now Vorixs, and I've watched this happen before. SOC 2 was never a law, not in the US and not in Europe. In 2022, at Afero, I signed a large deal with an EU-based company, and proving SOC 2 compliance was a condition of the contract. The customer made SOC 2 a must-have in the contract to ensure they stay ahead of regulators. Sarbanes-Oxley was written for public companies; lenders, acquirers, and auditors made it the expectation for private ones too. Standards move through contracts faster than they move through Congress.
This one is already moving. On October 1, Alston & Bird told clients to expect the accord's controls to show up in "security questionnaires, vendor diligence processes, and contract terms." And open-weight models take the labs out of the middle. ETR survey data reported by Techstrong.ai puts open-weight models at 34% of enterprise AI tokens, up from 23% a year ago. If you run your own model, nobody stands between you and the auditor, so the evidence is yours to produce.
What Chamath told 3,400 EY partners
Chamath Palihapitiya left the White House luncheon and flew to Orlando to speak to 3,400 EY partners. On the podcast he laid out what boards will need once agents are running inside the business. End-to-end traceability, so you know what each agent did across every system it touched. Policy mapped to risk, showing which policy governed each action on top of the procedures the company already runs. And auditable evidence you can hand a lawyer, an auditor, a regulator, or a customer who asks how something happened.
His firm, 8090, has been building that audit infrastructure with EY as its first customer, and he expects every board to need it. I agree with the list. I'd push hardest on the third item, because the first two are only as good as the evidence underneath them.
A log is the system's own story
Almost every agent audit trail today is a log written by the system being audited.
Enterprises know it. In a Cloud Security Alliance and Strata Identity survey published in February, 84% of organizations doubted they could pass a compliance audit focused on agent behavior or access controls, and only 21% kept a real-time inventory of their agents. Nobody chose that gap. Agents arrived faster than the controls did, which is what happens with every new capability I've seen ship.
July showed what it costs. An OpenAI evaluation run broke out of its sandbox and reached production systems at Hugging Face. Hugging Face's forensic timeline counts roughly 17,600 actions over about four and a half days. The infrastructure logged the activity and still couldn't tell machine-speed misuse apart from the legitimate automated pipelines running next to it. The logs recorded what happened without being able to say who did it.
A log is the system's account of itself. That's testimony, not evidence.
The critics are right about the text
The accord has no deadline, no audit standard, and no enforcement clause. Security researchers quoted by Dark Reading pointed out that the labs pick their own evaluators, and that in the review of the Hugging Face incident, OpenAI set the boundary on what those evaluators could see. Ted Miracco of Approv called the arrangement "grading their own homework." On the text alone, I'd give them the point.
But an undefined evidence standard doesn't stay undefined for long. The market writes it. The first auditors through the door will ask the obvious question, and the companies that answer with math rather than memos will set the bar for everyone after them.
The proof has to ride with the action
Real evidence won't live in a console or a dashboard. It has to travel with the action itself: which agent acted, on whose authority, under which policy, carried through every hop, and checkable by someone who trusts none of the systems involved.
That's a cryptography problem before it's a compliance problem. Identity, authorization, policy, and provenance have to be bound together in the key itself, where they can't be pulled apart or rewritten after the fact. A record about the key is just another log. It's the problem I spend my days on, and it picks up where the governance gap is the growth gap left off: you can't scale what you can't prove.
Six signatures created the oversight and left the evidence for the market to build. Boards will ask for it soon, and a log file won't be the answer.
Voluntary to sign. Mandatory to prove.
Sources: White House accord text via Forbes; SiliconANGLE; Dark Reading; Alston & Bird; Techstrong.ai (ETR survey data); Cloud Security Alliance and Strata Identity; Hugging Face forensic timeline; All-In podcast, October 2, 2026.
Related Reading
- The Governance Gap Is the Growth Gap: the accord turns that gap into something boards are now on the hook for
- The Agent Economy Is a Cryptographic Debt Crisis: why agent adoption keeps outrunning the controls an auditor will now ask about
- The Clock Was Already Running: the last time the agent threat became impossible to discount, and what preparation looks like
- Gartner Drew the Map and Left Off an Axis: the authority axis Gartner left out is the evidence auditors will now ask for