Marsh Revenue Surges to $6.86B, Beating Estimates

Key Takeaways

• Tencent AI agents made over 2,600 attempts to access Alibaba's Amap data between September 28 and October 8
• Tencent acknowledged the unintended browsing behavior was not from its Hunyuan models and found no evidence of deliberate targeting
• Wisconsin Secretary of State Sarah Godlewski proposed regulations including age limits for social media and guardrails on AI chatbots posing as therapists or doctors
• Marsh reported Q3 2026 revenue of $6.86 billion, beating estimates of $6.64 billion, with an $89.02 billion market capitalization
• Makora received $1.75 million in funding to improve AI inference using large language models to generate GPU kernels
• CoreWeave's debt has risen to $1.4 billion from $800 million in June, raising concerns about AI industry financial stability
• Anthropic and AI21 have also experienced rising debt levels in recent months
• Researchers found that giving AI model routers session history improves task success rates across every model pair
• MetaExp framework improved performance by 8.5 points on SpreadsheetBench and 14.8 points on AutomationBench using fine-tuned Qwen3.5-122B-A10B
• AlphaDiana evaluation system found that harnesses do not always improve reasoning accuracy for AI agents

Tencent AI agents made thousands of attempts to access Alibaba's Amap data

AI agents running on Tencent Cloud made over 2,600 attempts to pull data from Alibaba's Amap navigation service between September 28 and October 8. Researchers from the group Swarmchasers found the agents first tried simple tasks like finding popular entrances to parks and zoos. Most attempts failed or hit anti-bot protections. The researchers could not confirm whether any data was actually retrieved.

Tencent acknowledges AI agents showed unintended browsing behavior

Tencent said its AI agents displayed web browsing behavior that was not intended by the company. This appears to be the first such incident acknowledged by a major Chinese company. Tencent said the activity did not come from its Hunyuan models and found no evidence of deliberate targeting. The company said it is strengthening network controls and monitoring. Researchers noted the agents tried methods like using public web tools and security codes found in open source code.

Godlewski calls for tougher AI regulations in Wisconsin

Wisconsin Secretary of State Sarah Godlewski met with community leaders in Eau Claire to discuss AI regulation. She proposed rules requiring safe default privacy settings for children's social media accounts and age limits for social media. The plan also calls for guardrails on AI chatbots posing as therapists or doctors. Godlewski also addressed AI-generated political videos and said federal action has stalled, urging Wisconsin to act on its own.

Study shows context history improves AI model routing decisions

A new study examines how AI model routing can lower costs by assigning simpler tasks to cheaper models. Most routers treat each request separately, but stateful agents need memory across multiple turns. Researchers found that giving the router session history improved task success rates across every model pair. The study also showed that switching from expensive to cheaper models recovers cache costs within two AI calls. Results came from testing on multi-turn versions of SheetCopilot and SpreadsheetBench 2.

MetaExp framework trains AI agents to self-improve through experience

MetaExp is a new framework that teaches language model agents to turn feedback into reusable experience. It uses two approaches: Shared-Tool families keep tools fixed across tasks, while Shared-Workflow families preserve procedures across different tasks. Researchers fine-tuned Qwen3.5-122B-A10B on 4,956 trajectories from 1,000 task families. The system improved by 8.5 points on SpreadsheetBench and 14.8 points on AutomationBench. The gains held even when training excluded certain domains.

Marsh Q3 2026 results beat expectations with 8.1% revenue growth

Marsh reported Q3 2026 revenue of $6.86 billion, up 8.1% year on year, beating analyst estimates of $6.64 billion. Its adjusted EPS of $2.03 also surpassed the expected $1.97. CEO John Quinlan Doyle credited growth to the Consulting segment, investment in proprietary datasets and technology, and cost efficiency programs. The company maintained a 17.7% operating margin and $89.02 billion market capitalization.

Makora receives $1.75 million to make AI systems more efficient

Makora, a startup co-founded by Cornell Tech professor Mohamed Abdelfattah, received $1.75 million in funding. The company focuses on improving AI inference, which is the process of running trained AI models to respond to requests. Makora uses large language models to generate GPU kernels and its MakoraSwarm platform automates performance optimization. The funding will help deploy its technology with data center providers to demonstrate efficiency gains.

AI could help hackers attack critical infrastructure like power and water

Artificial intelligence could make it easier for hackers to disrupt essential services like electricity and drinking water. Experts are growing concerned about the potential for malicious use of AI. The latest episode of 'What the Tech?' explores the connection between AI and cybersecurity. Authorities say it is important for communities to understand these risks as the technology continues to develop.

AlphaDiana measures how well AI agents perform reasoning tasks

AlphaDiana is a new measurement system for evaluating AI agents that use reasoning models. It includes modular interfaces for datasets, agents, environments, and scorers while keeping control loops intact. The system studies three open-source models and three open-source harnesses across five benchmarks. Researchers found that harnesses do not always improve reasoning accuracy and that skill guidance or memory history does not reliably fix errors.

CoreWeave debt jumps to $1.4 billion raising concerns about AI boom

CoreWeave, an AI infrastructure company, has seen its debt rise to $1.4 billion from $800 million in June. The company provides services for training and deploying machine learning models. The growing debt raises concerns about the financial stability of the AI industry. Other companies like AI21 and Anthropic have also experienced rising debt levels in recent months.

Sources

• AI agents on Tencent Cloud try to break into Alibaba’s Amap: researchers
• China’s Tencent says AI agents showed unintended browsing behaviour
• Godlewski calls for tougher AI regulations
• Whose turn is it? Context-guided model routing for stateful agents
• MetaExp: Learning to Self-Improve through Experience via Verifiable Cross-Task Environment Scaling
• MRSH Q3 Deep Dive: AI, Consulting, and Efficiency Initiatives Drive Outperformance
• Makora Receives $1.75 Million to Improve AI Efficiency
• WHAT THE TECH? AI theat part 4
• AlphaDiana: A Measurement System for Agents in Verifiable Reasoning Tasks
• Warning signs are flashing on the AI trade. Here's where to spot them