A voluntary AI pact meets always-on agents
The White House accord promises safety reviews as OpenAI launches dots. The questions are who checks the safeguards, what becomes public and who can intervene.
A voluntary AI safety accord and a new generation of always-on agents arrived on the same day. Together, they raise a practical question: who checks that the safeguards work, and who can intervene when they do not?
On September 29, President Trump announced an agreement with leading technology companies, including OpenAI, Anthropic, Google and Nvidia. AP reports that the commitments include internal controls, outside audits and board-level review. These are voluntary undertakings. The agreement leaves room for future legislation; it does not establish that mandatory regime today.
Promises and public authority
The policy debate is still open. Roll Call reports that Senator Mark Warner sought unanimous consent on Tuesday for legislation creating an AI Safety Board within the Commerce Department. The proposal would give the board access to new models before release and require incident reporting. Senator Ted Cruz objected, criticizing the concentration of authority. That stopped the unanimous-consent request; it was not a final vote rejecting the bill.
The difference matters. A voluntary assessment can produce useful evidence, while mandatory oversight can establish obligations that companies cannot simply choose to avoid. Neither label alone tells us whether an assessment is rigorous. Readers should ask who chooses the evaluators, what they can inspect, which findings become public and what happens when a system fails a test.
Our view: an announcement is a starting point for scrutiny. Meaningful safeguards need evidence, consequences and a credible way to stop unsafe work. A list of prominent signatories cannot substitute for those things.
What this means for your next agent
OpenAI introduced dots on September 29: agents with their own cloud computers and connections to apps. Access initially covers eligible Pro and Business Premium users, with an administrator-enabled beta for Enterprise. Availability depends on market and plan.
OpenAI distinguishes background research from assigned actions. Its proactive research uses connected tools in read-only mode; assigned work can involve actions governed by instructions, Custom Rules and approval requirements. That distinction is worth checking before connecting accounts. The company acknowledges that the system is imperfect.
Start with a bounded task and the smallest set of connections it needs. Decide in advance which actions require your review, especially anything that spends money, sends a message or changes a shared file. Check the result against the original material before relying on it. These are our practical recommendations, not evidence that any particular agent is safe.
For the wider debate, keep the standards separate: useful features deserve evaluation, company commitments deserve scrutiny, and public decisions deserve public accountability. You can value a tool while demanding stronger safeguards around its development and deployment.