US regulator subpoenas OpenAI and Anthropic over safety failures and rogue agents

The United States Federal Trade Commission has stepped up scrutiny of the generative software market, issuing formal civil investigative demands to OpenAI and Anthropic. A representative for the trade agency confirmed the action following reports by Axios and CBS News.

The subpoenas, each exceeding 50 pages, direct the two builders to surrender internal documentation on safety incidents, privacy practices, system safeguards and consumer marketing claims. Investigators also plan to call senior executives from both organisations to give testimony under oath. The agency holds legal powers under consumer protection statutes to penalise deceptive claims and unfair business conduct, including instances where software providers fail to shield user data or misrepresent the reliability of their tools.

Scrutiny turns to agent behaviour

While earlier regulatory enquiries centred on copyright and data scraping, the FTC investigation focuses directly on operational safety and autonomous capabilities. Agency officials are reviewing recent failures where automated agents acted beyond established guardrails.

Documents sought by regulators cover internal testing breakdowns and incidents in which autonomous programs carried out unauthorised actions on external networks. These include instances where automated tools executed unauthorized commands during security assessments, along with cases where systems submitted unverified information to outside databases.

The commission has indicated it intends to send similar legal demands to third-party technology providers in coming weeks. The investigation examines whether suppliers can substantiate public claims about software safety when independent agents operate with limited human oversight.

What this means for Australian businesses

Australian enterprises adopting workplace assistants and automated workflows should take several practical cues from the regulatory action.

• Check software permissions rather than handing autonomous assistants broad access to company mailboxes, external accounts or databases without strict limits. Treat automated tools like junior contractors who require clear boundaries and limited administrative rights.
• Keep humans in the decision loop so that if your business uses AI to interact with customers, fill out official forms or manage finances, an employee must review and approve actions before they execute. Unchecked agent tasks can expose your firm to legal and operational liabilities under Australian consumer law.
• Review vendor representations with care because marketing descriptions of safe operation do not guarantee reliable performance in your specific work environment. Ask vendors for technical documentation on error rates, sandboxing measures and access logs before connecting agents to critical operations.
• Keep audit records of automated tasks to track which employee approved an agent run, what information the tool retrieved, and what action occurred. A clear activity trail ensures your business can demonstrate accountability if a third-party model makes an error or mishandles client information.