Microsoft's Project Perception Monitors AI Agents

Key Takeaways

• Microsoft's Project Perception uses Guardian Agents to monitor and control overprivileged AI agents inside enterprises
• Google DeepMind opened its Argon AI model after it solved 12 simulated cyberattack scenarios and helped find a critical bug
• Nvidia-backed doubleAI raised $220 million and claims its Maxwell model cut customer compute costs by 67 percent
• Mobileye CEO Amnon Shashua co-founded doubleAI, which plans to charge customers a share of savings achieved
• FIS launched an AML Investigations Agent for U.S. banks and an Agentic Insights Suite for CFO teams
• Exa launched Snapshot, a dated web search tool that lets AI agents backtest using historical indexes
• Carnegie Mellon University offers a free AI certificate for African learners, with applications closing November 15, 2026
• GS Securities integrated Acuity AI trading signals into its portal and MT5 platform for clients
• AdaptClose raised an undisclosed pre-seed round for its AI sales coaching platform targeting home services businesses
• Patna's state government plans to require AI training for teachers at state universities and colleges

Microsoft addresses risks from overprivileged AI agents inside companies

Microsoft warns that AI agents used inside companies are becoming powerful and hard to control. These agents often have too many permissions and can make decisions in important business processes without oversight. Microsoft calls them agentic insiders and says they need constant monitoring. The company proposes Guardian Agents to watch every action and reduce risks in real time. Project Perception is a security system that uses signals from identity, data, cloud, endpoints, network, and runtime to manage these agents.

Microsoft strengthens tools to manage AI insider threats in enterprises

Microsoft is improving its security tools to handle risks from AI agents working inside organizations. Project Perception is an agentic security system that uses a cyber stack to watch and control AI agents. It relies on Guardian Agents to moderate actions before they happen. The system protects identity, permissions, data, endpoints, network, cloud, and runtime layers. Enterprise AI agents can be unpredictable, overprivileged, and able to act in critical processes without human judgment.

Microsoft reveals plan to secure AI agents in businesses

Microsoft executive Hayete Gallot outlined a framework for handling AI agents inside enterprises. The company says these internal agents, called agentic insiders, may act unpredictably and often have too much access to data and systems. Microsoft is using Project Perception, first announced in July, to monitor agent activity with red, blue, and green agents. Guardian Agents supervise actions before they occur. The system provides traceability so security teams can review behavior and adjust permissions. Microsoft also says future security will need more transparency into model reasoning, including chain-of-thought tokens.

Carnegie Mellon University launches free AI certificate for African learners

Carnegie Mellon University is offering a free online AI certificate for Africa. The program includes six courses and takes about 45 hours per course to complete. It covers Python for AI, machine learning, generative AI, and AI in Africa. Applications are open from October 15 to November 15, 2026, and the next cohort starts January 17, 2027. Learners need basic knowledge of linear algebra, probability, statistics, and programming. The program is cost-free for qualified applicants and awards digital badges and a certificate.

Carnegie Mellon University offers free AI certificate for African professionals

Carnegie Mellon University is providing a free Artificial Intelligence certificate for learners across Africa. The program is open to students, graduates, and working professionals. Applications for the January 2027 cohort must be submitted by November 15, 2026. The certificate aims to strengthen AI skills and improve career opportunities. The program is free and designed to help learners build practical knowledge in artificial intelligence.

AdaptClose raises pre-seed funds for AI sales coaching

AdaptClose has exited stealth mode after raising an undisclosed pre-seed funding round. The startup created a platform that uses machine learning to provide personalized sales coaching for home services businesses. This tool helps companies of all sizes improve their sales performance and increase revenue. With this new funding, the company plans to expand its platform and boost its marketing efforts to become a leader in the industry.

Mexican auto dealers adopt AI for better sales and inventory

The automotive industry in Mexico is changing as dealerships increasingly use artificial intelligence technologies. A recent study shows that AI use in this sector is expected to grow by 25% in the next year. Dealerships are using these tools to analyze customer data, optimize inventory levels, and improve the overall customer experience. While there are challenges like the need for training and potential job changes, the benefits of data-driven decisions are clear.

Google opens Argon AI model after successful security trial

Google DeepMind has opened its Argon AI model to a broader rollout following a gated trial period. During the trial, the model solved 12 simulated cyberattack scenarios and helped security firm Wiz find a critical bug in hospital software. To ensure safety, DeepMind uses layered defenses including input screening and internal monitoring to prevent misuse and prompt injection. The company reports that these measures significantly reduced violations and successful attacks compared to previous models.

GS Securities adds Acuity AI trading signals to its platform

Dubai-based broker GS Securities has started using AI-powered trading signals from Acuity Trading. The partnership includes Trade Intelligence, Event Intelligence, and Market Intelligence to help clients understand market movements. Clients can access these insights through the GS Securities portal and MT5 desktop platform to see entry levels, stop-losses, and reasoning behind trade ideas. This integration provides traders with context and analysis to support their own decision-making.

doubleAI model cuts customer compute costs by 67 percent

Nvidia-backed startup doubleAI claims its new Maxwell model reduced an early customer's compute costs by 67 percent. Co-founded by Mobileye CEO Amnon Shashua, the company uses a code execution model that works inside running programs to improve hardware efficiency. The startup plans to charge customers a share of the savings they achieve, with a broader release coming soon. DoubleAI has raised $220 million and used half of those funds to buy 1,000 Nvidia B300 chips.

State universities to require AI training for teachers

The state government in Patna plans to make AI training mandatory for teachers at state universities and constituent colleges. Teachers will learn to use AI for classroom teaching, lecture planning, question paper preparation, and evaluating answer sheets. The move supports a broader AI policy for universities. The government also stressed the need for safeguards around accuracy, data privacy, academic integrity, and human oversight in AI-based assessment.

Exa launches dated web search tool for AI agents

Exa introduced Snapshot, a tool that lets AI agents search the web using dated indexes. This allows teams to backtest agent features without using future information. Exa uses its own crawling infrastructure, database, embedding model, and retrieval architecture across about 100 billion documents and trillions of URLs. The company serves hedge funds, asset managers, sovereign wealth funds, and VCs. The demo raised questions about pricing, data retention depth, and whether the versioned histories will meet compliance audit standards.

FIS launches AI tools for banks and finance teams

Fidelity National Information Services released its AI-driven AML Investigations Agent for U.S. banks. It also launched the FIS Agentic Insights Suite for CFO teams. Both tools are designed to connect directly with existing banking and finance workflows. The launches focus on high-value areas like financial crime reviews and cash and risk forecasting. These tools aim to improve efficiency, reduce costs, and support better decision-making.

Sources

• Securing the agentic insider
• Microsoft bolsters security to better control AI insider risks
• Microsoft outlines security framework for AI agents in enterprise
• The Carnegie Mellon University Artificial Intelligence Certificate for Africa (Free-online course)
• Carnegie Mellon University Offers Free Artificial Intelligence Certificate for Africa: Apply for the January 2027 Cohort
• Dealroom.co | AdaptClose exits stealth with pre-seed round to bring AI sales coaching to home services
• Mexico’s Automotive Dealers Enter the AI Era
• Google opens Argon after gated defender trial
• GS Securities implements Acuity AI trading signals
• Nvidia-backed doubleAI says its AI model cut an early customer's compute costs 67%
• AI training for state university teachers to be mandatory soon
• Exa lets agents search the web as it was yesterday
• What Fidelity National Information Services' AI Finance Tools Launch Means For Shareholders