Today’s top line: A governance and security probe at Nvidia tops the agenda while OpenAI’s reported >$10B quarter and Google launches a purpose-built Gemini Decisions API reshape industry economics and product choices.
Nvidia board hires Wachtell after CEO affair linked to suspected spy
Summary: Nvidia’s board has hired law firm Wachtell, Lipton, Rosen & Katz to conduct an internal investigation after CEO Jensen Huang admitted to an extramarital affair with Alice Huang, whom U.S. intelligence agencies believe was spying for China. Nvidia’s stock fell about 9.5% over two trading days following the revelations; the board has said it has “no reason to believe” secrets were compromised.
Digest: The probe raises corporate-governance and national-security questions for a company central to AI hardware. Watch for the investigation’s findings on what was disclosed to the board and any regulatory or shareholder actions that follow.
Source: TechCrunch
OpenAI’s quarterly revenue reportedly tops $10 billion
Summary: Bloomberg, via TechCrunch, reports OpenAI’s quarterly revenue crossed $10 billion for the first time, up from $6.7 billion in the second quarter. The company’s adjusted operating margin improved from -150% to -109% quarter over quarter and OpenAI’s annualised revenue run rate reached $50 billion, according to the report.
Digest: Strong demand and tighter pricing appear to be boosting OpenAI’s economics, shifting industry competitive dynamics and enterprise buying. Expect increased scrutiny of how vendors price models and shape enterprise contracts.
Source: TechCrunch
Google launches Gemini Decisions API for fast, calibrated multimodal choices
Summary: Google introduced the Gemini Decisions API powered by the gemini-decision-1 model, which scores closed-set choices (yes/no, multiple choice, rubric scores) in a single forward pass across text, images, audio, video and PDFs and can evaluate up to 1,000 independent questions per request. Google says the API returns probabilities and confidence scores in about 150-250 ms end-to-end and bills input tokens at $0.10 per 1M with zero output tokens charged.
Digest: Purpose-built decision APIs change how companies route queries, enforce policy checks, or grade content by returning calibrated confidences quickly and cheaply. Expect this to be used for model routing, content moderation and tool selection in production systems.
Source: Google
Grok Build link-sharing flaw exposed confidential apps and SpaceXSI files
Summary: Security firm Claroc found that Grok Build’s randomly generated link-sharing scheme made thousands of live projects enumerable and exposed sensitive data from many users. Claroc discovered data from two dozen SpaceXSI apps, including a confidential onboarding plan and a sales territory map covering 594 accounts; SpaceXSI removed the exposed links after being notified.
Digest: The incident shows how developer-facing, AI-enabled app platforms can leak sensitive corporate material through misconfigured sharing. Businesses deploying such tools should audit default sharing settings and verify access controls before use.
Source: Ars Technica
DeepMind uses AI to formally verify parts of the Signal protocol
Summary: Google DeepMind used a new AI framework, AlphaProof Code, to produce a formal verification of selected portions of the Signal encryption protocol in a four-week proof-of-concept project. The work found no new vulnerabilities in the checked portions, correctly flagged known older vulnerabilities in earlier Signal versions, and DeepMind released the specification and proof to help others.
Digest: AI-aided formal verification could speed rigorous security checks on critical code, but scaling and cost remain open questions. Companies that run high-stakes services should watch whether these tools become practical for supply-chain or protocol audits.
Source: Ars Technica
BotPsy: a 0.8B-parameter on-device assistant claiming top agentic scores for its size
Summary: Tether AI Research released BotPsy, a 0.8 billion-parameter agentic language model built on Qwen3.5-0.8B and designed for on-device personal-assistant tasks and secure workflows. The Hugging Face post says BotPsy outperforms every tested model up to 1.2B parameters on several tool-use and agentic benchmarks, with examples like BFCL v4 (44.0 vs 23.4 for its base) and ACEBench (58.7).
Digest: Smaller, agentic models tuned for tool use make private, on-device assistants more feasible and cheaper to run. Businesses building sensitive assistants should consider small guarded models for edge deployment, while recognising limits for high-stakes tasks.
Source: Hugging Face
EarthRecord benchmark reveals agents struggle to map pixels to real land change
Summary: The EarthRecord paper introduces a benchmark built on four decades of raw satellite archives across 56 sites and poses 238 land-change questions. Evaluations of twelve multimodal agents show most complete end-to-end analyses but many fail when the answer requires translating imagery to on-the-ground change; when one instrument’s data were altered, half of the correct answers followed that alteration.
Digest: EarthRecord highlights that current agents lean on data quirks rather than robust pixel-to-change understanding. Organisations using agents for monitoring or compliance should validate model outputs against independent references and be cautious in production decisions.
Source: arXiv
Quick hits
- Amazon debuts Ring Rover and Ring Drone security devices – Amazon announced a $1,000 Ring Rover robot (shipping from Dec 16, 2026) and a $250 Ring Drone (aiming for Jan 27, 2027), raising fresh privacy concerns about mobile home and business cameras. (Ars Technica)
- Databricks upgrades Genie agent for extended, unsupervised multistep work – The update lets the agent run longer multistep tasks using context from customers’ Databricks stores; CEO Ali Ghodsi demonstrated internal use cases. (TechCrunch)
- Google launches Tabletop for collaborative group planning – A Google Labs experiment that creates shared, editable spaces and lets a built-in prompt assistant generate templates like trip itineraries and polls. (Google)
- Plurai raises $13M to simulate and harden enterprise AI agents – Startup uses synthetic digital twins of company systems to surface edge cases and post-train smaller models for cost and reliability gains. (TechCrunch)