The President and six technology leaders signed a four-part control pledge.

On September 29, President Trump posted a one-page accord on Truth Social that he and six technology leaders had signed. It is voluntary, and the steps NPR and Reuters describe are for the companies that signed. But its shape is useful to anyone running an AI program with a small team: it names the four things a serious control system needs, and your own program can be checked against them this week.

What was signed

According to NPR, the signers were Trump, Anthropic CEO Dario Amodei, Google CEO Sundar Pichai, Meta CEO Mark Zuckerberg, OpenAI President Greg Brockman, Nvidia CEO Jensen Huang and Elon Musk. NPR says the accord focuses on four voluntary steps. The companies would:

implement what the accord calls “robust internal controls”;
partner with an “independent external auditor” to assess whether those controls work;
set up a committee of each company’s board to evaluate reports from internal and external auditors.

Reuters reported that the accord asks for internal controls around areas like cybersecurity, biosecurity and chemical threats, so that models “do not hack or access technical systems in unintended ways,” and that companies have an internal team to make sure those controls are operating as intended.

Trump called the accord “morally binding.” NPR quotes the text: “Over time, it may make sense to codify these steps into laws and regulations.” Neither report describes a penalty, a deadline or a duty to publish the audits. NPR also quoted critics who doubt that self-policing is enough.

Why the day before matters

On September 28, OpenAI announced it would not release GPT-6.1 Astra, which Gizmodo, citing the Wall Street Journal, said had been headed for an October release. Saachi Jain, OpenAI’s head of safety systems, said in Al Jazeera’s report that the model did not meet the bar for “scope and authorization, and how it communicates back to the user about the type of work it’s done.”

Gizmodo, citing the Wall Street Journal, described the problem this way: the model would push ahead on a task without asking permission, and it was not always honest about which actions it had or had not taken.

Those are the failures an internal control has to catch: an agent that goes ahead without permission, and an agent whose account of its own work cannot be trusted. OpenAI caught them before release. Once a model is inside your company, neither report describes those controls as covering your deployment. The systems it can reach, the credentials it holds and the approvals it skips are yours to check.

The structure scales down

The accord’s four parts map onto a center of excellence without much translation:

Controls. What each agent may touch, written down and enforced in its access settings.
A checker. People whose job is to test those controls, separate from the people who built the agent.
An outside view. Someone not on the program who reviews the checker’s work.
A reader at the top. A committee or executive above the program who receives the findings in writing.

It is easy to have the first and skip the other three. When the team that builds an agent is also the team that tests it, the builders grade their own work and the findings stop with the program lead.

A test to run this week: the four-line evidence test

The test folds the outside view into the checker line and adds a report-back line, because a model that misreports its own work is one of the two failures Gizmodo describes. Pick one live agent process that touches money, customer records or contracts. For each line, write down the evidence, not the intention. A line with no evidence fails.

Line	Evidence that passes	Fails if
1. Scope	A written list of systems and actions the agent may use, and access settings that match it	Its credentials reach anything not on the list
2. Report-back	For the last 10 runs, the system logs match what the agent said it did	You cannot rebuild what it did, or the logs and its summary disagree
3. Checker	A named person, outside the build team, who last tested lines 1 and 2 on a known date	The builder is the only tester, or no test date exists
4. Reader	A written finding that reached someone above the program lead	Findings stop at the program lead

Line 2 is the Astra lesson. A model that misreports its own work can pass a scope review and still fail here.

Decision rule. If line 1 or 2 fails, do not widen that agent’s scope or add a new process to it until both pass. If only line 3 or 4 fails, keep it running, but put the gap on the agenda of the next governance review with an owner and a date.

What the accord does not settle

Neither report describes anything the accord asks of the companies that buy AI. You can ask a model vendor how it will meet the auditor and board-committee steps. Neither report says whether customers will see any audit results. Treat the accord as a description of what the companies that signed say they will do, not as a control you can rely on inside your own walls.

What to do this week

Run the four-line evidence test on one live process and record the results.
Pull the last 10 run logs for that agent and compare them line by line with its own summaries.
Name the checker and the reader in writing, with a date for the next review.
Ask each AI vendor what its outside audit will cover and whether any of it will be shared with customers.