Google Gemini Decisions API Scores 1000 Questions Per Request

Key Takeaways

• Protegrity’s Cedar-based policy engine applies controls across data, permissions, prompts and responses; its AI Team Edition runs on Microsoft Azure and AWS, with Google Cloud and on-premises support planned.
• Obsidian added runtime checks for OpenAI Codex and governance for agents in Salesforce Agentforce, Databricks, Microsoft 365 and other enterprise platforms; Orca expanded discovery of AI-generated apps and agents.
• Commvault and Zscaler plan to link AI security controls with data protection and recovery, but have not announced a joint-product launch date. AgentRecover early access is targeted for Microsoft Foundry and Amazon Bedrock AgentCore.
• Google’s Gemini Decisions API scores up to 1,000 multimodal questions per request, supports inputs up to 256K tokens and costs $0.10 per million input tokens.
• In BuySideFlow tests, agents scored 16.2%–36.8% on results; 50.8% of runs that passed result checks still breached point-in-time rules.
• The NYSE’s proposed tokenized-securities venue would use Avalanche ledger technology and requires regulatory approval. The article also reports AlphaPepe fundraising, holder, partnership and launch figures.
• McKinsey reports global bank profits rose 7% to $1.3 trillion; its survey found 23% use generative AI for financial matters, while 72% of AI users would delegate at least one task to an agent.
• Jefferies sees Cohu’s potential serviceable market rising from $3.0 billion to $5.4 billion and initiated Corning coverage at Buy with a $134 target, citing AI data-center demand.
• AMD announced its 6th Gen EPYC server SP7 processors. Oracle is developing AI-assisted application-security tools, while self-hosted model operators are advised to use admission controls and manage retries and memory.
• The United States and China plan talks on AI and AI safety as Trump and Xi prepare to meet.

Protegrity adds policy engine to protect enterprise AI

Protegrity has added a unified policy engine, built on the open-source Cedar policy language, to control what users, AI agents and queries can access and do. Its controls use Protegrity’s personally identifiable information model to protect sensitive data, and apply across data, permissions, prompts and responses. The company also introduced Data & AI Fast Lanes and AI Gateway in the latest AI Team Edition, which runs on Kubernetes in Microsoft Azure and Amazon Web Services. Secure Query and Iceberg Protector are in technology preview, and Protegrity says customers are already using them. The company plans to add Google Cloud and on-premises support, including Red Hat OpenShift.

Obsidian adds runtime controls for AI agent actions

Obsidian has expanded its AI agent security tools with a Runtime Policy Decision Engine that checks agent actions against centrally managed policies. It uses identity, employment, application and customer-provided context, such as data labels, to decide whether to allow an action, block it or require human approval where supported. A new MCP Security Proxy checks agent tool calls before they reach business applications, while a Policy Decision API extends the controls to customer-operated gateways and custom tools. Runtime enforcement for coding agents now includes OpenAI Codex, alongside Claude Code, Cowork and Microsoft Copilot. Obsidian also added discovery and governance for agents built into Salesforce Agentforce, Databricks, GitHub, Microsoft 365, Snowflake, ServiceNow and Workday.

Orca expands tools to find AI apps, agents and code risks

On November 11, 2026, Orca Security announced expanded tools for finding risks from AI-generated applications, autonomous agents and code. Its AI AppGen Security features detect digital fingerprints left by tools such as Lovable, v0, Replit, Base44 and Bolt across cloud infrastructure, DNS records, workloads, code repositories and cloud identities. Orca connects these clues to permissions and resources to help teams assess whether an app can reach sensitive company data, and offers checks such as detecting disabled row-level security in Supabase projects. The company is also expanding support to Base44, OpenAI and Vercel. Its expanded AI agent security features provide visibility into cloud-managed and self-hosted agents, including coding agents running on cloud infrastructure.

Commvault and Zscaler plan AI security and recovery partnership

Commvault and Zscaler announced a strategic partnership on November 10 to connect Zscaler’s AI security controls with Commvault’s data protection and recovery technology. The planned work aims to help enterprises find AI tools, set policies for AI activity and agents, and restore affected data, applications and identities after problems. The announcement does not give a launch date for a joint product. Separately, Commvault’s AgentRecover is designed to map what agents interact with and help teams understand changes and plan recovery; early access is targeted for Microsoft Foundry later in November and Amazon Bedrock AgentCore in December. Those timelines apply to AgentRecover, not the Zscaler partnership.

Jefferies keeps Cohu Buy rating as AI demand could boost growth

Jefferies reiterated its Buy rating and $90.00 price target for Cohu after the company’s Analyst Day strengthened the firm’s confidence in its AI opportunity. Analyst Kevin Garrigan said rising device complexity, power needs and test times could lift Cohu’s serviceable market from $3.0 billion to $5.4 billion. He said qualified high-performance computing programs support the near-term Eclipse ramp. Jefferies sees a path to $1.3 billion in revenue, $5 in earnings per share and an 18% free cash flow margin. Cohu shares had gained 88% over the past year and were at $43.57 when the article reported the analysis.

Jefferies Sees Corning’s AI Data Center Opportunity Growing by 2028

Jefferies began covering Corning with a Buy rating and a $134 price target, which suggests 19% potential upside from Tuesday’s close. The firm said demand from AI data centers and multi-year contracts with large cloud companies and phone carriers could support Corning’s fiber business. It expects the Optical business to scale up in 2028, while the Photonics business may grow more slowly. Progress in next-generation pluggable optics and co-packaged optics could help the stock. Corning reported third-quarter adjusted earnings of $0.91 per share and revenue of $5.08 billion, up about 19% from a year earlier.

Google Launches Gemini Decisions API for Fast Multimodal Scoring

Google introduced the Gemini Decisions API, powered by its Gemini Decision 1 model. The API can evaluate up to 1,000 questions in one request across text, images, audio, video, and PDFs, with inputs of up to 256K tokens. It scores yes or no questions, multiple-choice options, and ordered rubrics, returning probabilities and confidence scores instead of generating answers one token at a time. Google says end-to-end latency is 150 to 250 milliseconds and charges $0.10 per 1 million input tokens, with no charge for output tokens. Developers can access an API key in Google AI Studio and use its Python, JavaScript, and REST guides.

BuySideFlow Benchmark Finds Many Investment Research Agents Break Time Rules

Researchers introduced BuySideFlow, a benchmark for testing whether large language model agents use financial information that was available at the time of an investment decision. It includes 408 practitioner-derived Chinese-language tasks covering stocks, funds, fixed income, and macroeconomics, based on a fixed snapshot of a licensed financial database. The researchers tested nine agent setups, which scored between 16.2% and 36.8% for result accuracy. An audit of 900 runs on 100 shared tasks found point-in-time violations in 193 of the 380 runs that passed result checks, or 50.8%. The study says matching an expected result does not prove that an agent used valid historical data.

NYSE Selects Avalanche for Tokenized Securities as AlphaPepe Promotes AI Trading Tools

The article says the NYSE plans a venue for round-the-clock trading and settlement of tokenized securities, with Avalanche providing ledger technology for its clearing firm. The platform still needs regulatory approval, and NYSE technology will match orders. It also promotes AlphaPepe’s AlphaSwap trading platform, which offers contract screening, whale tracking, and trend signals, and says AlphaSwap v2.0 is due later this week with a redesigned trading flow, security patches, and bug fixes. The article lists ALPE at $0.03647, says Stage 20 has sold out and Stage 21 is live, and gives a planned DEX launch in Q1 2027. It also reports more than $3.35 million raised, over 13,000 holders, five announced CEX partnerships including Coinstore, and completed audits from BlockSAFU and Coinsult.

Swiss Banks Lead on Profits but Face Growing AI Challenges

A McKinsey study says global banks raised profits by 7% to $1.3 trillion last year, while revenues before risk costs rose 3% to $6.4 trillion. Swiss banks performed especially well, with record assets under management and stock market valuations above European and other international peers. The study says digital banking use in Switzerland is two to three percentage points above the European average, but large international banks often have more money and stronger technology systems for AI. In a survey of about 46,000 customers in 13 countries, 23% said they use generative AI for financial matters, and 74% of those users rely on outside platforms such as ChatGPT, Claude, or Gemini. Still, 62% of AI users would prefer financial AI services from their current bank, and 72% would delegate at least one financial task to an AI agent, usually with approval before a transaction.

AMD releases 6th Gen EPYC processors for data centers

AMD announced the availability of its 6th Gen EPYC server SP7 processors. The announcement comes as advances in AI chips continue to reshape data center computing. The article describes the processors as part of ongoing developments in data center technology.

Oracle develops AI tools to investigate application security threats

Oracle is developing AI-assisted application security tools to connect application, identity, and infrastructure signals with business context. The system aims to identify unauthorized activity while allowing legitimate work to continue. It uses security models and specialized AI agents to examine requests and behavior across sessions, test possible explanations, and guide verified fixes. Oracle says security teams should check which evidence sources are available and treat missing data as a coverage limit, not proof that an application is safe.

Self-hosted AI models need controls to manage demand

Harshul Jain and Parv Khatri told AI Engineer that self-hosted large language models need admission controls to handle bursts of up to 40 agent calls. During inference, the prefill phase builds a KV cache and the decode phase generates tokens, while memory stays reserved until a request finishes or disconnects. At the point requests enter the system, operators can admit them with status 200, defer them with 503 and a retry-after value, or reject them with 429 for caller overuse and 529 for pool overload. Agents should begin timeouts when the first token arrives, close dead streams to free cache, and spread retries using jitter.

Trump and Xi plan talks on AI risks amid US-China rivalry

The United States and China plan to discuss artificial intelligence and AI safety. The talks are expected as President Trump and Chinese leader Xi prepare to meet. The discussions come amid global competition for technology leadership.

Sources

• Protegrity Adds Unified Policy Engine to Secure Enterprise AI Data and Agents
• Obsidian expands AI agent security with runtime policy enforcement
• Orca Security adds visibility into AI-generated code and agent risks
• Commvault and Zscaler Partner on AI Security and Recovery
• Jefferies reiterates Cohu stock Buy rating on AI opportunity
• GLW Stock Could Ride The AI Data-Center Buildout, Jefferies Says: Sees Corning’s Optical Business Scaling In 2028
• Gemini Decisions API: Fast, calibrated multimodal decisions
• BuySideFlow: A Point-in-Time Benchmark for LLM Agents in Investment Research
• Latest Crypto News: NYSE Picks Avalanche for Tokenized Stocks as AlphaPepe Builds Its AI Trading Platform Before Token Listing
• Swiss banks especially profitable, but pressure is growing in AI
• Latest News In AI Chips - Advancing Data Centers with AMD's New EPYC Processors
• From signals to action: AI-assisted application security investigations
• Self-Hosted LLMs Must Learn to Say No
• Trump, Xi to discuss AI risks as US-China tech rivalry intensifies