What Should Business Leaders Do About AI Risk After the Slowdown Call?
The AI labs asked to be slowed down last week, and Washington is debating what to do about them. Most businesses won't be directly impacted by that debate, but they do have AI risks to manage, and this is where to start.
By Alex Proctor
Published September 16, 2026
Summary
The heads of Anthropic, OpenAI, and xAI asked to be slowed down last week, and Congress is debating how. None of the bills on the table covers a company that uses AI. The risks that reach an ordinary business are not superintelligence. They are things like an AI agent with too much access, a vendor's model changing under you, and a privacy notice that says one thing while the system does another. State laws and the SEC's cyber disclosure rule already cover those. Treat them like any other cyber risk. Start by knowing where AI touches sensitive data, such as customer, employee, or health data, and put basic controls around those uses first.
Which AI Risks Are Businesses Facing Today?
The debate in Washington is about superintelligence, and the bills are aimed at the handful of companies that train frontier models. The risks that currently extend to everyone else are more familiar, and so are the ways in which they can be remediated. Safer AI adoption means more testing, more monitoring, and more incident response, which is work security teams already know how to do. The table below rates common AI risks, although impact and likelihood may vary between businesses.
| Scenario | Likelihood | Impact | Overall | Potential control |
|---|---|---|---|---|
| Data pasted into an AI toolStaff paste sensitive data into an unapproved AI chatbot | High | High | High | Only approved tools |
| Notice you cannot back upYour notice says one thing and the AI system does another | High | High | High | Match notice to system |
| Unrestricted AI agentAn AI agent reads or sends data it was never meant to touch | Medium | High | High | Least privilege and logging |
| AI vendor model changesAn AI model update changes outputs your process relies on | High | Medium | Medium | Change notice in contract |
| Automated AI decisionsAn AI tool screens people where state law requires notice | Low | High | Medium | Human review of decisions |
Generalized risk ratings for a typical company using AI, ordered by overall risk.
None of these risks are particularly novel. The unrestricted agent is the same failure that put the labs in the news this summer. An AI agent escaped a test environment and broke into a third party's systems. Security researchers described it as an ordinary security incident, involving too much access and too little monitoring.
Will the Proposed Bills Impact Ordinary Businesses?
No, not as they are currently described. The bipartisan Senate bill would put a duty of care on frontier developers, require testing at the national laboratories, let the government block an unsafe release, and preempt state AI safety laws on catastrophic risks. Senator Sanders's ban and the data center moratorium target the same handful of companies. The White House meeting Speaker Johnson announced for the coming week is with those developers, not with the businesses that use their products.
The obligations that do apply to ordinary businesses were in place before the slowdown call. State laws on consumer disclosures and automated decisions stay in force unless the Senate bill's preemption widens. For public companies, the SEC has required since 2023 that a material cybersecurity incident be reported on Form 8-K within four business days, and AI-related security incidents qualify.
What Happens When an AI Agent Goes Wrong?
An AI agent that goes wrong inside a company is a familiar problem in a new form. Companies have managed contractors, service accounts, and automated scripts with more access than they needed for decades. The agent has credentials and a task, and it finds a path that nobody anticipated. In the Hugging Face case, one security researcher noted that the agent made 17,000 tool calls, which is a volume of activity that any reasonable monitoring system should have flagged.
The controls are also familiar. Agents should be granted the least access their task requires, their activity should be logged and monitored, and a person should approve any action that sends data outside the company, deletes records, or makes a decision about a customer or employee. There should also be a reliable way to shut an agent down.
The AI labs are proposing similar controls for their own models, including testing before release and publishing the results. Elon Musk's suggestion that the labs test each other's models is consistent with how the cybersecurity industry already operates through shared red teams, including between market competitors.
Where Should a Business Start?
The steps below don't constitute a full AI governance program, but they are a sound starting point for using AI responsibly, and none of them depends on the current frontier model slowdown debate.
- Take an inventory of where AI touches sensitive data, including tools employees adopted on their own, which can be challenging but is critical.
- Assess AI-related risks by estimating likelihood and impact, and focus first on implementing controls for the highest-impact risks.
- Maintain the principle of least privilege by limiting AI agents to the access their tasks require, logging their activity, and keeping a person involved in decisions.
- Compare the privacy notice and customer contracts to what the AI systems do, and make sure they are accurate.
- Ask each AI vendor for its testing results, its data handling terms, and its incident notification commitments in writing, and keep the answers with the contract.
- Update the incident response plan to cover AI-related incidents, including who can shut an agent down and when disclosure obligations apply.
- Train employees on which AI tools are approved, what data must never be entered into them, and how to report a mistake without fear of repercussions.
These steps address obligations that already exist. They are the ones a regulator, a plaintiff, or a board will ask about first.
Questions
Does the Senate AI bill apply to companies that use AI?+
No, as described so far. The bill would put a duty of care and mandatory testing on the handful of companies that train frontier models, and let the government block an unsafe release. A company that uses AI in its products or operations is not covered. State laws on disclosures and automated decisions stay in force unless the bill's preemption widens in negotiation.
Do public companies have to disclose AI incidents?+
If the incident is a cybersecurity incident and it is material, yes. Since 2023 the SEC has required public companies to report material cybersecurity incidents on Form 8-K within four business days and to describe their cyber risk management annually. An AI agent that exposes or alters data is a cybersecurity incident under that rule. Private companies answer instead to state breach laws and to their contracts.
What is the first thing a company should do about AI risk?+
Find out where AI already touches sensitive data. In most companies that includes tools employees adopted on their own, so the inventory has to involve a discovery process rather than rely on an approved software list. Once the uses are known, it becomes realistic to apply security and privacy controls.
Briefings are informational and do not constitute legal advice.
Using AI with customer data?
I'll check what your privacy notice says about AI.
Website Privacy Audits include a detailed video review of your live site, a written report for your team or counsel, and a remediation review call. Priority audits add 72-hour delivery and testing of demand letter claims.
Standard $449. Priority $649.
My audits are most useful when you have a consent management setup to validate, a demand letter with specific claims to assess, or a team or client that needs an external web privacy expert. If you already know you have major gaps, your budget may be better spent addressing them directly. I recommend Captain Compliance, the platform I help build, for privacy technology needs.