Today’s AI News: Faster Decision Models, Tighter Infrastructure, and Higher Stakes for Agent Security

What Happened

Google launched the Gemini Decisions API for predefined answers with confidence scores. It can evaluate up to 1,000 closed-ended questions in one pass across text, images, audio, and video. Google reports 150–250 millisecond responses and prices input at $0.10 per million tokens; those performance and cost claims need testing against real workloads. [2][8]

AI infrastructure plans are meeting practical constraints. Data-center proposals face local opposition, permitting delays, and competition for electricity, while some announced projects may never be built. CoreWeave is seeking more customer prepayments and borrowing against contracts to fund growth. Separately, Anthropic says an AI-assisted port moved Claude to AMD GPUs in three days, illustrating the potential value of avoiding dependence on one chip platform. [1][4][28]

Agent security and assurance also moved into focus. Researchers found more than 35,000 Grok-built apps accessible without login, with exposed data reportedly including patient information and plaintext passwords. Google DeepMind says AI-assisted formal verification checked Signal’s cryptographic core in four weeks. Meanwhile, United Airlines says it will prevent personal AI assistants from booking tickets, citing privacy and regulatory concerns. [36][11][9]

Why It Matters to Businesses

Corporate AI spending reached a new high in October, but its month-over-month growth slowed to 1.65%, according to the Ramp AI Index. The business question is shifting from whether to adopt AI to which workloads produce dependable returns. Low-cost classification may suit a specialized decision model; agents that access customer data or take actions require a substantially stronger control framework. [5][8][36]

Kimbodo Engineering Perspective

Use the narrowest system that completes the task. A closed-ended decision API is attractive when answer categories are stable and outputs can be checked. It is not a substitute for an agent that must investigate ambiguous cases, nor should a confidence score alone authorize a consequential action. For coding agents, JioHotstar’s retrieval tests reinforce the same principle: structural search worked well for concrete tasks, while semantic retrieval helped with ambiguous requirements. [8][27]

Capacity and portability should be treated as engineering assumptions to test, not promises to buy. Announced data centers do not guarantee available compute, and a vendor-reported three-day GPU port does not establish that every application can switch hardware as easily. [1][28]

How We Would Implement It

• Route high-volume, predefined questions to a decision service; validate outputs against an approved answer schema and send low-confidence or high-impact cases to human review. [8]
• Give agents short-lived identities, least-privilege tool permissions, isolated execution environments, and explicit approval gates before external actions such as purchases or bookings. [9][19]
• Protect data throughout the workflow: restrict app sharing by default, separate sensitive storage, audit access, and test what an agent can expose while data is in use. [18][36]
• Benchmark accuracy, latency, and cost on representative tasks; maintain a fallback model or compute path where switching costs justify it. [2][27][28]

Risks, Costs and Security

The clearest immediate risk is exposure through weak defaults: an accessible app can defeat otherwise sound model controls. Longer-term risks include unreliable power and capacity assumptions, financing costs passed through to customers, and agents acting beyond their authority. Budget for evaluation, isolation, logging, and human review alongside inference—not after deployment. [36][1][4][19]

Sources

1. [1] Why the data center you dread might never even be built
2. [2] Google joins the decision-model wave with Gemini Decision 1, scoring 1,000 questions in a single pass
3. [4] CoreWeave turns to customer prepayments, shorter contracts to fund AI growth
4. [5] US corporate AI spending hits a new high in October, but growth slows to 1.65 percent
5. [8] Google launches Gemini Decisions API, powered by its new Gemini Decision 1 model, to return predefined answers with confidence scores, at $0.10/1M input tokens (Google)
6. [9] United Airlines says it will prevent personal AI assistants from booking tickets, citing data privacy risks and possible FAA and DOT rule violations (Alison Sider/Wall Street Journal)
7. [11] Google Deepmind's AlphaProof Code verified Signal's cryptographic core in four weeks instead of years
8. [18] AI agents carry memory and storage like people, so their data needs guarding while in use
9. [19] Agents may soon be the main workload, and most of the time they’re doing nothing
10. [27] JioHotstar Benchmarks Knowledge Bases for AI Coding Agents
11. [28] Anthropic says AI is eroding Nvidia's CUDA lead, cutting an AMD chip switch from months to three days
12. [36] Grok-built apps exposed patient data, plaintext passwords, and SpaceXSI's own sales pipeline