US regulator probes OpenAI and Anthropic over AI agent safety claims

The US Federal Trade Commission (FTC) has opened a wide-ranging investigation into how AI companies handle the risks of "agentic" AI, according to a daily AI roundup published on 1 October. The tools in question don't just answer questions. They take actions on your behalf.

What happened

The probe reportedly targets OpenAI, Anthropic and METR, an outside group that assesses AI models. The question is whether incidents involving AI agents, and the way safety is described to customers, could breach existing US consumer-protection law.

The Wall Street Journal reportedly expects the FTC to send formal information demands, seek testimony from executives and look closely at third-party assessors. These details come from a secondary roundup of the WSJ report, so treat them as early reporting rather than confirmed outcomes.

Notably, the FTC isn't waiting for new AI-specific legislation. It is testing whether rules already on the books apply.

The same day, the White House reportedly took a softer approach. A group of major AI companies signed a voluntary safety accord covering internal evaluations, outside audits and board reviews. The accord says it could be made law later.

Why it matters in Australia

This is a US investigation, and it doesn't directly change anything for Australian businesses. But it points to a trend that's worth noting if you're using AI tools or considering them.

• Agents are now the main event. More tools can send emails, book things, handle payments and update records without a person approving each step. That's useful, but it raises the stakes when something goes wrong.
• Safety claims are being tested. If regulators are checking whether vendors' safety promises hold up, it's sensible for you to do the same before you hand over real responsibilities.
• Existing laws can apply. Australia has also leaned on current laws, such as consumer and privacy law, while broader AI settings are debated. Your obligations don't pause because AI is involved.

Practical steps

• Start agents on low-risk tasks, such as drafting and sorting, before letting them act on their own.
• Keep a human approval step for anything involving money, customer data or external messages.
• Ask vendors where your data goes, what logs are kept and what happens if the agent makes a mistake.
• Keep a simple register of which AI tools your team uses and what they can access.
• Make sure someone in your business owns AI oversight, even if it's only part of their role.

The bottom line

AI vendors are facing more scrutiny of what they promise about agents. For Australian businesses, the lesson is to adopt agents carefully, with clear limits and records. We'll keep watching how this develops and whether similar scrutiny follows locally.