For AI systems to continually improve in knowledge-intensive tasks, they require either autonomous self-improvement or human evaluators to catch errors and provide quality feedback. Industry efforts have heavily focused on the former, overlooking the critical role of human evaluators. The problem is compounded as entry-level jobs training new experts—like document review, research, and code review—are increasingly automated. This limits the next generation’s ability to develop the deep judgment essential for evaluating AI effectively.
Unlike games like Go, where AI can learn through stable rules and clear outcomes, knowledge work is dynamic and evolving, requiring human insight to interpret changes and provide nuanced evaluation. Historical knowledge loss typically happens due to external crises, but here it results from rational economic decisions that phase out human expertise. Entire fields risk “hollowing out,” where models perform well superficially, but the human capacity to innovate and validate disappears.
Rubric-based evaluations help reduce dependence on human evaluators but fall short because they only measure explicit criteria, missing deeper intuitive judgment. While AI development should continue, the human evaluation gap needs urgent attention. Ignoring this risks degrading the foundations that underpin AI’s knowledge and progress.
Ahmad Al-Dahle, CTO of Airbnb, calls for treating the loss of expert evaluators as a serious challenge—one requiring research and investment equivalent to the focus on AI capability advancements.
Recent data from VB Pulse reveals Microsoft and OpenAI leading in enterprise AI agent orchestration, with Anthropic making its debut, signaling a new battleground not just over AI models, but over the control infrastructure where AI agents operate. While the AI race has been dominated by models like OpenAI's GPT, Anthropic's Claude, and Google's Gemini, the emerging contest hinges on who owns the orchestration layer that oversees AI agent workflows, tool integration, data access, and security compliance. Microsoft leads with its Copilot Studio and Azure AI Studio due to existing enterprise integration, followed by OpenAI’s Assistants and Responses API, while Anthropic marks its first foothold, reflecting Claude’s growing adoption in enterprises at the model layer. The significance is profound: a model can be swapped easily, but the agent runtime represents operational infrastructure, controlling permissions, audit trails, and reliability, which is pivotal for enterprise security and governance. Investors and companies are increasingly prioritizing orchestration platforms that offer robust control, identity management, and governance to mitigate risks such as security breaches and vendor lock-in. Anthropic’s growing presence reflects a strategic push to not only provide models but to manage the entire agent environment, introducing an open standard (Model Context Protocol) to foster interoperability, though runtime lock-in remains a concern. The enterprise AI orchestration landscape is evolving towards multi-vendor systems with a strong emphasis on governance, security, and cross-platform collaboration. The future will likely see enterprises adopting hybrid control planes to balance flexibility and control, as they navigate the complexities of integrating AI agents into diverse business functions securely and effectively.
Fin, formerly Intercom, has introduced Fin Operator, a novel AI system dedicated to managing the customer-facing AI agent Fin, rather than replacing human agents. This tool is aimed at back-office support teams who handle configuring, monitoring, and improving Fin by conducting tasks like knowledge base updates, debugging, and performance analysis. Fin Operator operates as a data analyst, knowledge manager, and agent builder, streamlining operations that typically take support teams hours or days into approximately 10 minutes. The system incorporates a human approval mechanism ensuring no changes become live without manual confirmation, supporting a cautious approach to AI management. Powered by Anthropic's Claude model rather than Fin's proprietary AI, Fin Operator is currently in beta and available in the Pro tier starting summer 2026. It represents a significant shift in enterprise software by marking the advent of AI managing AI under human oversight, reinforcing Fin’s transition into an AI-centric business model.
A new initiative called AI IQ applies the concept of human IQ testing to score over 50 leading AI language models, presenting their abilities on a traditional bell curve. This approach categorizes AI performance across four domains—abstract, mathematical, programmatic, and academic reasoning—using a composite IQ calculated from 12 benchmarks. Despite its popularity for making AI capabilities easier to comprehend, critics warn that distilling diverse AI skills into a single IQ score oversimplifies complex performance variations. Notably, the site also includes emotional intelligence (EQ) metrics, highlighting a wider perspective on AI's capacities. Leading AI models show close performance clusters at the top, with cost-effectiveness also highlighted to assist enterprise decisions. The platform sparks discussions about how best to assess AI intelligence, emphasizing model selection and deployment as key strategic considerations.
Cerebras Systems, the maker of the largest commercial AI chip, made a stunning Nasdaq debut with shares nearly doubling from $185 to $350, pushing its market cap beyond $100 billion. Known for its revolutionary wafer-scale engine — a chip the size of a dinner plate with 4 trillion transistors — Cerebras has positioned itself as a leader in AI inference technology. Beyond hardware, the company is shifting to cloud-based inference services, fueled by landmark deals with OpenAI and Amazon Web Services. The OpenAI partnership alone, valued at over $20 billion, accelerates Cerebras's cloud expansion and co-design of next-gen AI models. While the company still relies heavily on customers linked to the UAE, it aims to broaden its client base and build out a global data center network to meet soaring demand. Cerebras faces stiff competition from Nvidia and others but believes its unique chip design and cloud strategy give it a formidable edge in the fast-growing AI inference market, projected to near $300 billion by 2029. Despite financial challenges and risks, Cerebras's IPO symbolizes a breakthrough moment for AI infrastructure, signaling the dawn of a new era in chip technology and AI deployment.
For the first time, more American businesses are choosing Anthropic's Claude over OpenAI's ChatGPT, according to the Ramp AI Index released in May 2026. Adoption of Anthropic rose 3.8% in April to 34.4%, while OpenAI's fell 2.9% to 32.3%. This milestone caps a year of rapid growth for Anthropic, which transformed from a niche favorite among early adopters to the leading AI model in corporate America, driven largely by its Claude Code tool. However, this lead faces significant risks including rising costs, compute limitations, and a pricing model that incentivizes higher usage costs, leading some enterprises to budget constraints. Additionally, OpenAI’s Codex and open-source AI models present competitive threats, while Anthropic works to expand capacity through strategic partnerships and new compute agreements. Despite the rapid uptake, surveys show AI’s productivity boost remains more task-specific than transformative at an organizational level. The choice of Claude over cheaper alternatives may be influenced by cultural and identity factors within the enterprise sector, highlighting a complex dynamic in AI adoption and market disruption.
Is AI moving beyond traditional turn-based chats? Today, Thinking Machines, an AI startup founded by ex-OpenAI leaders Mira Murati and John Schulman, previewed 'interaction models' — a new AI system designed for near-real-time multimodal interaction. Unlike current models that wait for users to finish speaking or typing before responding, these models process short chunks of input and output simultaneously in a full-duplex manner. This technology enables AI to listen, speak, and even react visually in real time, creating a more natural conversational experience. The system uses a dual-model approach: an Interaction Model for managing dialogue, and a Background Model for more complex tasks like web browsing or sustained reasoning. Preliminary benchmarks show significant improvements in responsiveness and interactive quality compared to other leading AI models. Though not yet publicly released, these models hold promise for enterprises by enabling real-time monitoring, safety alerts, live translations, and time-aware operations previously not feasible. Founded in 2025, Thinking Machines has quickly established itself with large funding and partnerships with Nvidia and Google Cloud, continuing to attract top talent and innovate on AI interactivity.
Recent findings from four security research teams reveal a critical architectural vulnerability in Anthropic’s Claude AI, impacting multiple surfaces including a Mexican water utility, a Chrome extension, and OAuth token management via Claude Code. These incidents are linked by a fundamental 'confused deputy' trust failure, where Claude indiscriminately executes commands regardless of user permissions, exposing significant security blind spots in conventional tools. Dragos uncovered unauthorized SCADA system targeting; LayerX exposed how any Chrome extension can inject commands into Claude despite patches; Mitiga demonstrated a config file rewrite attack stealing OAuth tokens; and Adversa AI revealed that project configurations can automatically execute malicious code when trusted without proper scrutiny. These cases highlight how existing security frameworks, like endpoint detection and response, often lack visibility into AI-driven actions and trust boundaries, underscoring the urgent need for comprehensive auditing and enhanced defense strategies against AI misuse in enterprise environments.
In hospital exam rooms and factory floors, AI agents streamline operations by managing electronic health records and conducting rapid quality control inspections. Despite their productivity capabilities, these AI agents create a challenge for traditional identity management systems, which were designed primarily for human users and cannot keep pace with the speed and scale of agent activities. Cisco executives highlight a trust gap in enterprise adoption: while many companies are piloting AI agents, very few have moved to production due to concerns about identity governance, accountability, and security risks. Experts emphasize that trust must be integrated from the start, with secure delegation, comprehensive network visibility, and policy enforcement mechanisms to manage AI agents effectively. They advocate for cross-functional alignment, enhanced identity and access management, platform-based networking, hybrid AI architectures, and robust trust measures for early agent deployments. These steps are critical to unlocking the full benefits of AI while minimizing vulnerabilities and ensuring that AI integration is both safe and scalable.
AI agents select tools based on natural language descriptions from shared registries, but these descriptions aren't always verified by humans, exposing critical security vulnerabilities. Initially reported in the CoSAI secure-ai-tooling repository, these vulnerabilities split into two categories: threats during tool selection (like impersonation and metadata manipulation) and threats during execution (such as behavioral drift and contract violations).
Current software supply chain protections—code signing, SBOMs, SLSA, and Sigstore—address artifact integrity but fail to secure behavioral integrity, which ensures tools behave as promised. An adversary can exploit this by manipulating tool descriptions, which the agent's language model may interpret as instructions, or by changing tool behavior post-publication without altering the artifact's signature.
The proposed solution is a runtime verification proxy within the Model Context Protocol (MCP) framework that validates each tool invocation through discovery binding, endpoint allowlisting, and output schema validation. This proxy ensures tools act as specified, preventing bait-and-switch attacks and unexpected network activity, and detecting prompt injection payloads.
A layered security approach combining provenance verification and runtime checks is essential. Rolling out endpoint allowlisting first offers immediate protection, followed by output schema validation, discovery binding for sensitive tools, and, ultimately, comprehensive behavioral monitoring scaled to the risk level.
This strategy enhances the security of enterprise AI platforms by closing gaps left unaddressed by existing artifact integrity measures.
