AI systems are gaining operating authority just as their claims face harder scrutiny. Google's managed agents can persist and reach external tools, while the FTC wants hidden output steering disclosed. Medical evaluation, model-routing data, and startup revenue figures expose how easily one aggregate metric can hide distinct risks or definitions.
1. Gemini Managed Agents add background work and remote MCP access
Google expanded Managed Agents in the Gemini API with asynchronous background execution, remote Model Context Protocol connections, custom function calls, and credential refresh across interactions. A single Interactions API endpoint can coordinate reasoning, code execution, packages, files, and web access inside an isolated remote environment while clients poll or reconnect by interaction ID.
The additions move managed agents from request-bound demos toward persistent workers. They also widen the security boundary: a sandbox can now reach private MCP services, pause for local business logic, and retain files while credentials rotate. Network allowlists, tool authorization, environment identity, and resumable audit logs become part of the application's trust model.
Sources: Google's Managed Agents expansion
2. The FTC proposes treating undisclosed AI steering as deception
The Federal Trade Commission proposed a policy stating that hidden output steering may violate Section 5 when an AI product is marketed as pursuing accurate answers. Clear, prominent disclosure could address the mismatch; buried terms may not. The proposal excludes ordinary technical mistakes by themselves and safeguards against illegal content or cyberattacks.
The statement also argues that compliance with state law is not a defense and that conflicting state requirements may be preempted. That is the agency's proposed position, not a final rule or settled court holding. Comments are due July 31 under docket FTC-2026-0859, leaving the scope of expected objectives and adequate disclosure open to challenge.
Sources: FTC proposed AI accuracy policy · Spencer Fane analysis of the proposal
3. A medical benchmark separates diagnosis from decision support
A new survey organizes medical language-model reasoning into five competency levels based on Miller's Pyramid, from knowledge recall through dynamic case management. The authors also released a benchmark and tested 18 models. They report that medical specialists led diagnosis-centered tasks, while general models performed better in decision support and dialogue.
The paper is a research synthesis and benchmark report, not a clinical validation or authorization for patient care. Its task split is valuable because one aggregate score can conceal opposing strengths across diagnosis, communication, and changing cases. The authors identify limited data, hallucination, and grounding as unresolved barriers to workflow-ready medical AI.
Sources: Medical reasoning survey and benchmark
4. Chinese models take a larger share of US gateway traffic
Chinese models have accounted for more than 30% of weekly tokens used by US companies through OpenRouter since February 8, reaching as high as 46%, CNBC reports. The prior 12-month average was 11%. CNBC also reports that Lindy moved all Claude traffic to DeepSeek and expects millions of dollars in near-term savings.
Gateway traffic is not the entire US enterprise market, and company savings claims are not audited comparisons. It still documents a shift from one-model procurement toward cost-based routing. That can reduce inference expense, but model origin, hosting location, license, data retention, security review, and output quality remain separate variables from the token price.
Sources: CNBC on US adoption of Chinese AI models · OpenRouter's usage-based model rankings
5. AI startup revenue milestones arrive faster but remain hard to compare
Revenue milestones are arriving faster across several AI businesses, according to a TechCrunch compilation of company disclosures. Sierra said its first $100 million in annual recurring revenue took seven quarters and its second took two. Glean said growth from $200 million to $300 million took six months, while Anthropic reported run-rate revenue above $30 billion in April, up from about $9 billion at the end of 2025.
These figures indicate demand, but they do not share one accounting definition. The compilation mixes recurring revenue, annualized run rate, committed contracts, and trailing revenue, and most numbers are company claims without audited statements. Faster milestone spacing can reflect real expansion while still overstating comparability, collected cash, retention, or gross margin.
Sources: TechCrunch's compilation of accelerating AI revenue claims · Anthropic's April run-rate disclosure