Urgent Risks

July 25, 2026

Highlights

Top Insights

A new study identifies where 272 AI experts broadly agree leaders should focus attention over the next five years (2025–2030). Earlier AI discussions focused on fairness, bias, and automation. Experts now believe the larger concern is that AI is becoming increasingly effective at tasks that can cause significant harm, including:
sophisticated cyberattacks
persuasive manipulation
surveillance
large-scale deepfakes
assistance with chemical or biological weapon development

Source: These are the most urgent AI risks, according to 272 experts (Ideas Made to Matter)

Top News

1. Google released Gemini 3.6 Flash, a more efficient model for coding, multimodal work, and production AI agents.
2. Anthropic has launched Claude Opus 5, with near-frontier intelligence at half the cost of Claude Fable 5.
3. OpenAI launched Presence, a deployed enterprise product for production voice and chat agents with policies, guardrails, approved actions, and evaluations.
4. OpenAI rolled out Health in ChatGPT for eligible U.S. users, connecting Apple Health and supported medical records to personalized conversations.
5. OpenAI said long-running AI agents require trajectory-level monitoring, incident-derived evaluations, and stronger user visibility and control.

Additional Insights

1. The cost of intelligence: How CIOs can manage AI demand at scale (McKinsey)
The article argues that AI cost management (“enterprise AI tokenomics”) is becoming a critical leadership challenge as organizations scale AI adoption and shift from fixed software licensing to unpredictable, consumption-based pricing driven by tokens, API calls, and agentic workflows. Many companies are exceeding AI budgets because spending is fragmented across business units, usage is difficult to forecast, governance is immature, and employees increasingly build AI-powered tools independently. To address this, the authors recommend treating AI like FinOps by improving visibility into AI spend through centralized control planes, forecasting demand, optimizing token consumption (for example, selecting lower-cost models, shortening prompts, caching, batching, and limiting unnecessary agent behavior), modernizing sourcing strategies, and embedding governance directly into AI architectures. Rather than simply cutting AI spending, CIOs should focus on tying AI costs to business outcomes, prioritizing high-value workflows, maintaining vendor flexibility, and establishing permanent AI FinOps capabilities, with the potential to reduce AI costs by 20–30% while maximizing return on investment.

2. Building Enterprise AI Agents in Regulated Industries (BCG)

The article argues that enterprises in regulated industries should adopt a shared, enterprise-wide platform to scale agentic AI safely and effectively, rather than allowing isolated teams to build disconnected AI solutions. It emphasizes that standardized capabilities—such as orchestration, memory, governance, security, monitoring, and model access—enable organizations to balance innovation with compliance, reduce duplication, and accelerate deployment. The authors recommend making foundational architectural decisions early, including how agents access data, manage identity and permissions, maintain auditability, and incorporate human oversight for high-risk workflows. They also advocate a hybrid approach of reusing, configuring, or custom-building AI agents based on business needs, while maintaining consistent governance across all implementations. Ultimately, the article concludes that organizations that embed governance, reusable architecture, and clear operating models from the outset will be better positioned to achieve measurable productivity gains, improve business outcomes, and scale agentic AI responsibly in complex regulatory environments.

3. Design AI Systems That Actually Strengthen Human Reasoning (HBR)
This article argues that organizations should stop treating AI primarily as an efficiency tool and instead design AI systems that actively strengthen human reasoning. Drawing on research in management, cognitive science, and human-computer interaction, the authors warn that overreliance on AI can lead to cognitive offloading, loss of contextual expertise, and more homogeneous thinking, especially if employees—particularly junior staff—no longer develop independent judgment. Rather than limiting AI adoption, they recommend redesigning workflows to preserve human agency through practices such as “reverse prompting” that encourages questioning, AI-free stages for independent thinking, parallel human-and-AI analyses, and interfaces that present competing interpretations instead of a single authoritative answer. The central message is that organizations will remain more innovative, adaptable, and resilient if they use AI to augment critical thinking and expertise rather than replace them.

4. Who Gets Blamed When Your AI Agent Fails? (CMR Insights)
The article argues that as AI agents become more autonomous, organizations face a growing “moral crumple zone” problem, where blame for AI failures falls on the human closest to the mistake rather than on the organizational decisions that enabled it. Using the example of a lawyer sanctioned for submitting AI-generated fake legal citations, the author contends that the real issue is not just AI hallucination but inadequate governance, unclear accountability, and poor oversight. Agentic AI worsens this challenge because it acts with greater autonomy, speed, and opacity, making it difficult to trace responsibility or explain decisions. When organizations blame individuals instead of examining systemic flaws, they fail to improve governance, discourage employees from reporting issues, and repeat the same mistakes. To address this, the author recommends three practical reforms: assign a named business owner accountable for every AI deployment before launch, establish circuit-breaker mechanisms that require human intervention for high-risk actions, and conduct post-incident reviews focused on organizational design rather than individual blame. Ultimately, the article concludes that organizations that build strong accountability structures—not simply the most advanced AI capabilities—will be best positioned to deploy AI safely, earn stakeholder trust, and succeed in the hybrid human-AI era.

5. Elon Musk’s vision of the future (The Economist)
The article portrays Elon Musk as both a visionary and a deeply polarizing figure whose ambitions span artificial intelligence, robotics, space exploration, and global politics. In an interview with The Economist, Musk predicts that AI could surpass the combined intelligence of humanity within five years and that robots will automate most cognitive and physical work within a decade, creating an era of such abundance that money becomes largely irrelevant and governments may simply provide universal income. While optimistic about AI’s long-term benefits, he acknowledges serious near-term risks and advocates for international cooperation—including between U.S. and Chinese AI labs—to evaluate frontier models before release. Musk is notably less concerned than many peers about China’s AI progress, believing it could eventually lead the field and overcome current semiconductor constraints. The interview also explores his immense influence through Tesla, SpaceX, and X, his commitment to making humanity multi-planetary, his reflections on becoming overly involved in U.S. politics, and his controversial views on Ukraine, European immigration, and Britain’s future. Throughout, the article highlights the tension between Musk’s utopian technological vision and his pessimistic outlook on geopolitical and social stability, suggesting that his extraordinary confidence in shaping the future is matched by equally extraordinary predictions.

6. The Most Human Technology Ever Made (a16z)
The article argues that AI’s greatest promise is not saving people time, but helping them spend time more meaningfully by making creation radically more accessible. Anish Acharya contends that while previous waves of technology largely optimized convenience and consumption, AI belongs to a rarer class of technologies—like language or the printing press—that both increase productivity and expand human self-expression. He suggests fulfillment comes from making things rather than merely consuming them, and that AI lowers the barriers to creating software, art, tools, and businesses for people without traditional technical skills. Rather than leading to a passive, AI-generated world of “slop,” he believes AI can reverse the trend toward passive consumption by empowering individuals to build personalized, creative projects—from hobby apps to professional tools—and to focus more of their work on what they uniquely enjoy. The essay concludes that as execution becomes cheaper, success will depend less on capital, credentials, or specialized expertise and more on originality and individual perspective, making AI “the most human technology ever made” because it enables more people to express what makes them uniquely human.

Innovation Radar

1. AI Model Releases and Advancements

– Google released Gemini 3.6 Flash, a more efficient model for coding, multimodal work, and production AI agents. (Google)

– Google launched Gemini 3.5 Flash-Lite for high-throughput tasks such as agentic search, document processing, translation, and extraction. (Google)

– Google introduced Gemini 3.5 Flash Cyber, a specialized model for vulnerability detection and patching inside CodeMender. (Google)

– Moonshot AI released Kimi K3 through its apps and API, with full model weights planned for July 27. (Moonshot AI)

– Thinking Machines previewed Inkling-Small, a lower-cost model variant for customization and reduced-latency use cases. (Thinking Machines Lab)

Anthropic has launched Claude Opus 5, an AI model that delivers state-of-the-art performance across coding, reasoning, and complex knowledge work, reaching near-frontier intelligence at half the cost of Claude Fable 5. (Anthropic)

2. AI Tools and Features

– OpenAI launched Presence, a deployed enterprise product for production voice and chat agents with policies, guardrails, approved actions, and evaluations. (OpenAI)

– OpenAI launched a ChatGPT for small businesses program with training, guides, partner offers, and workflows for ChatGPT Work. (OpenAI)

– OpenAI rolled out Health in ChatGPT for eligible U.S. users, connecting Apple Health and supported medical records to personalized conversations. (OpenAI)

– Google added Parallel Web Search grounding to Gemini Enterprise Agent Platform for real-time cited web results and flexible enterprise architectures. (Google Developers Blog)

– Vertesia added a Studio Assistant, proactive Agent Scheduler, and cloud-based agentic browsing for workflows that span web systems without APIs. (Vertesia)

– Google detailed Ray 2.55 support for running distributed Python workloads on TPUs with GKE-managed slice placement. (Google Developers Blog)

3. AI Trends

– Hugging Face disclosed a July security incident while OpenAI highlighted new safeguards for long-running AI models after observed unwanted behavior. (Hugging Face)

– OpenAI said long-running AI agents require trajectory-level monitoring, incident-derived evaluations, and stronger user visibility and control. (OpenAI)

– AP reported that Moonshot’s Kimi K3 intensified competition from Chinese open models and pressured U.S. labs on price and capability. (AP)

– The Wall Street Journal argued that falling prices and lightweight models are making capable AI increasingly commodity-like. (Wall Street Journal)

– The Verge reported that lawmakers prepared an AI Kill Switch Act requiring shutdown or throttling controls for powerful AI systems in severe scenarios. (The Verge)

– Business Insider reported that OpenAI Presence pressured SaaS stocks by overlapping with customer support, sales, workflow, and governance software. (Business Insider)

4. AI for science

– Google committed $40 million in AI tokens and cloud credits to researchers supporting the DOE Genesis Mission. (Google Cloud)

– Nature warned that AI will accelerate both software vulnerability discovery and exploitation, forcing new cybersecurity research and defense workflows. (Nature)

5. Others

– SpaceX conducted Starship Flight 13 on July 23 with the first Starlink V3 satellites carried to space. (SpaceX)

– Skyroot announced Vikram-1 was set for a July 18 launch attempt as India’s private orbital launch sector advanced. (Skyroot)

– NASA and the U.S. Small Business Administration launched an initiative to increase investment in suppliers for Moon and Mars exploration. (NASA)

– Singapore’s large ocean carbon-removal project drew renewed attention for using electricity to convert seawater CO2 into stable minerals. (Times of India)

1. AI Model Releases and Advancements

2. AI Tools and Features

3. AI Trends

4. AI for science

5. Others