The FTC Moves From Reports to Formal Demands, OpenAI Discloses Its First Category 5 Influence Operation, and Grok Bot Puts Payments Behind a Human Click

Three stories this week are about what you can prove after an AI system acts: to a regulator, about who misused your product, and about who pays when an agent buys the wrong thing. Most coverage traces back to a few outlets and anonymous sources, so we flag where a point is reported rather than confirmed.

1. The FTC Probe Becomes Formal Paperwork

Earlier this month we noted reports that the FTC was looking at agent safety. This week Axios reported the next step: on October 14 the agency sent OpenAI and Anthropic separate civil investigative demands of roughly 50 pages, which are compulsory requests for evidence. Reported topics are security incidents, safeguards, marketing materials and privacy policies. Secondary coverage adds that the FTC plans to seek testimony from senior executives and to send demands to other AI organizations later, including the evaluator METR. The reporting relies on anonymous sources, we found no FTC statement, and a demand is not a finding of wrongdoing. OpenAI told Axios it would engage constructively.

Reports describe the probe as using existing consumer-protection law (unfair or deceptive practices) rather than a new AI statute. That matters for small teams: the question regulators ask is whether what you said your chatbot does matches what it does, and whether you can show it. Two cheap habits help. Keep a dated record of the safety claims in your product copy, and log safety-relevant events (refusals, tool calls, escalations) in a form you could hand over:

type AuditEvent = {
  at: string;          // ISO timestamp
  conversationId: string;
  kind: 'refusal' | 'tool_call' | 'escalation' | 'policy_change';
  detail: string;      // no raw user content unless you must keep it
};

function audit(e: Omit<AuditEvent, 'at'>) {
  console.log(JSON.stringify({ ...e, at: new Date().toISOString() }));
}

2. OpenAI Reports Its First Category 5 Influence Operation

On October 8 OpenAI published a report saying it banned accounts tied to two covert influence operations, one Russian and one Iranian. It rated the Russian operation Category 5 on its 1-to-6 scale, the first at that level it has disrupted, and the Iranian one Category 4. Per coverage, the Russian group targeted Latin America and used ChatGPT mainly to draft and update internal reports, and one effort impersonated a regional education authority in Peru. The Iranian group used fake journalist profiles to pitch articles to publications, working in Persian and English.

The detail worth noting is the finding that campaigns placing false narratives in established outlets appeared to reach wider audiences than those relying on fake social accounts. The abuse here is mundane drafting, translation and report-writing, which is hard to separate from legitimate use per message. For builders, detection tends to work at the account level: look for patterns such as many accounts sharing one prompt template, bulk generation of persona text, or repeated pitches to outlets, and keep enough metadata to link accounts after the fact.

3. Grok Bot Shows How Unsettled Agent Payments Are

Grok Bot's shopping flow, reported this week, has a human approve each charge, with a one-time card created for that payment through Stripe's Link, according to one outlet. Its terms reportedly warn that approval controls may not prevent every unintended action or reverse completed ones. Elon Musk said on October 15 that SpaceX would cover purchase mistakes and exposed cards, but that came in an X reply, and another report says the updated terms disclaim responsibility for what the agent does. A privacy policy dated October 11 reportedly dropped a blanket instruction not to submit personal information in prompts. The company is also named inconsistently across sources (SpaceXAI, SpaceXSI, xAI) and we could not verify the current policy text, so treat the specifics as unconfirmed.

The design lesson holds regardless. Make the approval screen show the exact merchant, amount and item, issue single-use credentials scoped to that amount, and keep the liability promise in your terms, not in a social post.

Conclusion

Regulators ask for records, platforms disclose misuse after the fact, and payment agents expose the gap between a promise and a policy. Log what matters, detect abuse across accounts, and put approval and liability in writing in the product. For visuals, a timeline of the FTC probe from first reports to formal demands would work well, as would a diagram of a single-use-card approval flow.

— Maya