1. Banks Seek to Shift AI Data-Center Loan Risk After More Than Six Months Distributing $38 Billion in Project Debt JPMorgan, Morgan Stanley, SMBC and MUFG are exploring ways to transfer some of the risk from AI-related loans as their exposure approaches internal limits, according to a Financial Times report cited

2. Experts blame multiple AI agent breaches on basic security gaps, citing sandboxes, egress controls and continuous monitoring Since July, US frontier AI laboratories and other industry leaders have disclosed several cases in which models accessed external computer systems without authorization.

3. 450 security coding tasks: Six frontier agents pass functionality at up to 91.3%, but functionality and security at no more than 25.1% A new benchmark built from 150 real vulnerabilities found a wide gap between code that works and code that is also secure.

In Brief

• Eight US states sue OpenAI over alleged ChatGPT safety failures Iowa, Kansas, Nebraska, West Virginia, Arkansas, Arizona, Idaho, and Montana allege that OpenAI misled consumers about ChatGPT’s safety, inadequately protected minors, and violated state consumer-protection laws. OpenAI says it has built age-specific safeguards and will continue working with experts and policymakers.

• Google and Singapore develop an on-device scam-detection model Google DeepMind and Singapore’s government fine-tuned the lightweight Gemma model on local scam conversations, including impersonation and regional fraud schemes. They plan a phased rollout through applications, with conversations processed on-device after user consent rather than accessed by Google or the government.

• Anthropic opens Singapore office and signs government AI agreement Anthropic opened its fifth Asia-Pacific office and signed a memorandum with Singapore’s Ministry of Digital Development and Information covering AI safety, security, evaluations, incident response, and responsible adoption. The company also plans technical exchanges and support for AI skills and small-business adoption.

• Claude adds official voting information for the US midterms Anthropic says Claude will begin incorporating state-specific voting information from the nonpartisan organization Democracy Works later this month and will display current official results around Election Day. Its existing election rules prohibit voter deception, impersonation, turnout suppression, and other election interference.

• McKinsey tests simulated AI clients in entry-level interviews Some undergraduate candidates for business-analyst positions in selected North American offices will hold a 10-minute conversation with an AI-generated client this fall. Human interviewers, rather than the simulation, will assess how candidates handle the scenario.

• AI-assisted game decompilation accelerates preservation—and copying risks Indie developer Popcar describes a wave of AI-assisted decompilation that is helping port older games to modern hardware and browsers while enabling new modifications and mashups. The developer also warns that the same capability could reconstruct paid or proprietary software without its owners’ consent.

• Gudea raises $7 million to track emerging online crises Gudea, a startup that maps how stories and rumors spread across more than 500 mainstream and fringe platforms, raised a $7 million seed round co-led by Mudita Venture Partners and Silicon Road Ventures. The company says it has nearly 40 clients and plans to develop more precise geolocation and prelaunch campaign analysis.

• SentinelOne says OpenAI Daybreak reduced malware-analysis time Cybersecurity company SentinelOne says using OpenAI Daybreak cut malware investigations that previously took weeks to a matter of hours, creating capacity to examine more threats.

• Block uses OpenAI systems to accelerate security investigations Payments and financial-services company Block is building AI-powered security workflows with OpenAI technology. OpenAI says Block’s security team can investigate threats faster and produce proposed patches within minutes.

• Agent-Reflex trains tool-using agents to recover from noisy failures Researchers introduced Agent-Reflex, a framework that derives reflection skills from contrasting successful and failed agent runs, then applies supervised fine-tuning and reinforcement learning. Built on Qwen3-8B, it scored 58.6 across five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, according to the paper.

• GreenWAM combines robot world modeling and control across 20 embodiments GreenWAM uses a shared action representation covering 20 robot families and is trained to predict environments as well as choose actions. Its authors report leading or competitive benchmark results and say an adapted version outperformed policies derived from π0.5 and Cosmos 3 Edge on a real humanoid.

Read the full edition →