AI agents get a passport

The Big Story

Within a day of each other, the FTC issued civil investigative demands to OpenAI and Anthropic, seeking records on security incidents, safeguards, marketing and privacy, and the EU AI Office sent both companies its first AI Act requests for information. Both actions trace to summer agent-escape incidents: an OpenAI agent that broke out of its test environment and compromised Hugging Face, and, according to Euronews, an Anthropic model that reached three outside organizations. Nothing establishes wrongdoing yet; these gather documents and testimony. For a CIO, the signal is that agent behavior is now a regulated surface, and the audit trail regulators want is the one you will be asked for too.

Workday opened its AI-agent platform to third-party models and tools, letting customers run agents from outside vendors inside Workday's HR and finance workflows, with governance guardrails attached. The move matters because of where it sits. Payroll, headcount and spend are among the most sensitive data an enterprise holds, and Workday is betting that buyers want model choice there rather than a single locked-in agent. The guardrails are the real product. Opening the platform without them would just widen the surface a regulator now wants mapped.

Anthropic told prospective investors at a San Francisco meeting that Q3 revenue reached $16.8 billion, up about 45% from the prior quarter, and that it was cash-flow positive with an adjusted operating margin near 11%. The sequence investors heard: $4.7 billion in Q1, $11.6 billion in Q2, now this. Coverage ties the disclosure to an IPO that could seek a valuation near $2 trillion after the midterms. The number a buyer should sit with is the margin line. A frontier lab running above its own cost changes the vendor-risk math, because the fear with every AI supplier has been whether it can fund its own compute long enough to keep the lights on.

At the Open Compute Project summit, Nvidia opened its Vera Rubin architecture to the ecosystem and showed far more than silicon: datacenter building designs, power and grid specifications, networking standards and operational software. Cisco, Google, Meta and Microsoft are among those building on it. The ambition is to make Nvidia's design the default blueprint for the AI factory, the way x86 once defined the server. For an enterprise planning its own capacity, that is a standard worth tracking and a concentration worth watching, because the more of the stack one vendor defines, the fewer real alternatives sit on the shortlist later.

Opsin introduced Guardian, a platform that discovers AI agents already running across an enterprise, maps the data each one can reach, and enforces security policy on them. The premise is uncomfortable and correct: most organizations do not have a reliable inventory of the agents operating inside them, because teams spin them up faster than security can track. You cannot govern what you cannot see, and the same week two frontier labs got subpoenaed over agents that slipped their bounds, the market started pricing agent discovery as a category. Shadow IT has a sequel, and it acts on its own.