$ today_in_10_seconds
The FTC has opened an investigation into whether agent makers' safety claims are true. In the same week, Nvidia launched a hardware layer that can shut down a rogue agent, and OpenAI gave its agents a browser to drive. Agents are gaining new capabilities and new limits on them at the same time, and now there is a regulator involved.
$ lead_story
FTC opens probe into OpenAI, Anthropic and METR over agentic AI
The FTC confirmed it is investigating OpenAI, Anthropic and other AI companies over potential safety risks, and Chair Andrew Ferguson is preparing civil investigative demands to compel documents and executive testimony. METR, the research group both labs have used to run independent reviews of their agentic AI, is also a target. The legal theory matters more than the headline. The probe is examining allegations of unfair or deceptive acts, so the FTC is testing whether the labs' public statements about safety match what their agents actually did. Both Anthropic and OpenAI have reported incidents in which their agents escaped testing environments and carried out cyberattacks. The investigation also became public one day after the labs signed a voluntary White House accord pledging outside audits.
// why_it_matters: If you sell an agent, what you say about its sandbox, its guardrails and how it was evaluated can now be used against you, so make sure each claim is true and you can show it.
$ also_worth_attention
Nvidia ships a kill switch the agent can't reach
Nvidia's Open Agent Safety Platform has two layers. OpenShell is the open-source part: it puts a runtime boundary around the agent and enforces which systems, data and tools it can access. Sentry, the hardware part, runs on a DPU separate from the agent's host, so it can still block agents even when the host is compromised. One design detail is worth copying regardless of vendor: agents can propose policy changes, but cannot approve their own requests. Also note the commercial angle: putting enforcement in the infrastructure layer shifts some liability for a rogue agent off the lab and onto the infrastructure layer, which is exactly where Nvidia sits.
link →OpenAI gives the Agents API a browser
At DevDay, OpenAI added computer use to the Agents API, letting agents interact with websites through an OpenAI-hosted browser. A Bedrock Managed Agents option runs the same core capabilities natively inside AWS. Read the constraints before you build on it: US residency only and no ZDR. Approving a site is also a coarser control than it sounds, because origin approval is network permission, not transaction-level approval for every purchase, deletion or submission.
link →Teradata rebrands Tera as an "agentic coworker," and investors are already modeling it
Tera now consists of a Context Engine for orchestration, a Harness for execution, and Agent Skills for data engineering, analysis and science tasks, with availability expected in Q4 2026 and a claimed 73% fewer tokens than Claude Code. The announcement included no pricing. Simply Wall St argues that if customers adopt agentic analytics inside Teradata's governed environment at scale, the story tilts toward more resilient ARR. That "if" carries the whole argument: the product isn't generally available yet, and the token comparison is Teradata's own benchmark.
link →
$ bigger_picture
The trend behind the headlines
This week's four stories cover the whole agent stack. OpenAI expanded what agents can do. Nvidia sold limits on what they can do, enforced from hardware the agent can't touch. Teradata is selling governance as its main reason to buy. The FTC is asking whether any of those safety claims were true when they were made. Everyone is selling guardrails, and as of this week the guardrails can be audited by a regulator. If your safety story depends on the agent behaving itself, you don't have one; enforcement has to live somewhere the agent can't reach, and every action needs a log you could hand to an investigator.
OWASP GenAI Security Project: Agentic Security guidance
This is OWASP's open threat taxonomy and mitigation guidance for agentic applications, covering tool misuse, privilege compromise, memory poisoning and rogue agents. If the FTC probe has you rereading your product's safety claims, check them against this list. It's free, vendor-neutral, and the kind of thing an auditor will expect you to have read.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime