OpenAI has launched a significant update to its Codex AI platform, designed to extend its use from developers to a broad range of business professionals. This update introduces "Sites," a feature for rapid, semi-private web hosting within Codex, allowing teams to create and share dynamic, interactive web pages that transform static data into functional applications. Additionally, the "Annotations" tool enables precise, localized edits within documents like spreadsheets, preventing unnecessary full-document changes and preserving formatting integrity.

To further support enterprise needs, OpenAI has rolled out role-specific plugins encompassing 62 popular business applications and over 110 automated skills, covering functions in data analytics, creative production, sales, product design, and investment banking. These integrations simplify complex workflows without requiring custom API development.

Targeted at non-developers such as financial analysts, marketers, and researchers, who are rapidly adopting Codex, the update marks a strategic move to establish the platform as a daily operational tool in white-collar environments. The new features are accessible through the Codex app and plugin directory, available primarily to Business and Enterprise users under OpenAI's proprietary licensing model, with continued access through existing subscription plans.

Microsoft has introduced Microsoft Execution Containers (MXC) at its annual Build developer conference, addressing a critical concern in AI agent deployment: security. MXC is embedded directly into Windows and the Windows Subsystem for Linux, enabling developers and IT administrators to define strict execution policies that the OS enforces at runtime, ensuring AI agents can only access authorized resources. This policy-driven sandbox supports various isolation levels — from lightweight process isolation to micro-VMs and cloud instances — guaranteeing flexible protection based on the agent's function and risk. Key features include session isolation to prevent UI spoofing, input injection, and data leakage, plus binding agents to unique identities for precise audit and control. Microsoft demonstrated MXC's strength with an AI agent prevented from deleting desktop files due to these enforced boundaries. The system integrates with Microsoft's enterprise security tools like Defender, Entra, Intune, and Purview under the Agent 365 umbrella, set for preview in July, allowing centralized governance and compliance. Major industry players like OpenAI, Nvidia, Manus, and Nous Research are early adopters, signaling MXC's potential as a new standard for secure AI agent operation on Windows. By embedding containment at the OS level, Microsoft offers a uniquely practical, scalable solution, positioning Windows as the trusted platform for autonomous AI in diverse environments.

Enterprises face a significant challenge as each new AI agent starts without memory of business operations, data locations, or rules, leading to data silos. At Microsoft Build 2026, Microsoft introduced solutions to tackle this: Microsoft IQ and Rayfin. Microsoft IQ unifies four critical context sources—Work IQ (daily organizational workflows), Foundry IQ (institutional knowledge and rules), Fabric IQ (live operational business data), and Web IQ (real-time external world data)—providing AI agents a comprehensive understanding of the enterprise environment. Meanwhile, Rayfin is an open-source SDK and CLI that deploys agent-built applications directly into Microsoft Fabric, ensuring all application data integrates into the same governed platform, avoiding silo creation. This approach creates a virtuous cycle where agents draw from a shared data ontology and contribute back to it, enhancing organizational knowledge continuously. While other platforms like Snowflake, Pinecone, and Redis also develop context-rich AI layers, Microsoft's integrated and governed system aims to simplify execution and build trust in enterprise AI deployment, moving beyond just model availability toward scalable, compliant AI at the core of business operations.

In early 2026, VentureBeat's Pulse Research exposed the “Governance Mirage” highlighting a gap between AI governance structures and actual control layers in enterprises. Nearly half identified central AI governance ownership, but many reported confusion or obstacles like vendor opacity. The critical failure in AI deployments isn’t the intelligence of models but the runtime infrastructure. Enterprises relying on stateless setups like Python scripts and LangChain face challenges: lost context after restarts, escalating costs, and compounding errors leading to failures, causing teams to spend excessive time on maintenance rather than innovation.

The survey of 132 tech leaders across industries reveals that treating runtime durability as a core engineering priority separates successful deployments from failures reminiscent of past RPA pitfalls. Key findings include significant engineering effort spent on «plumbing» systems, a costly observability burden especially on Microsoft platforms, and a fragmented migration away from stateless architectures toward durable, multi-layered orchestration approaches. Firms seek a balance between AI autonomy and human oversight, with user acceptance rates becoming a critical metric for production readiness.

Security remains a challenge with no dominant strategy, as enterprises adopt varied approaches like policy-as-code and sandboxing to protect sensitive data. The shift toward architectures combining flexibility with durability—the “polyglot” approach—is emerging as a frontrunner, reflecting firms’ desire for control beyond vendor lock-in. The report underscores that while models are increasingly capable, the runtime environment and economic factors largely dictate AI’s enterprise viability today.

Perplexity AI, a rapidly growing search startup valued at $20 billion, showcased its groundbreaking hybrid local-cloud inference orchestrator at Computex 2026. This system dynamically directs AI workloads—some processed locally on devices using Intel's Core Ultra Series 3, others routed to advanced cloud models—balancing privacy, accuracy, and efficiency in real time. CEO Aravind Srinivas and Intel CEO Lip-Bu Tan demonstrated how sensitive data remains on the device while complex analysis proceeds in the cloud, eliminating the need for users to preselect task execution locations. This technology promises to reduce cloud costs, enhance latency, and meet strict data governance requirements, particularly for regulated enterprises such as finance and healthcare. Alongside industry moves by Nvidia and Intel unveiling new AI-optimized chips, Perplexity's approach could shift how AI workloads are managed, empowering more on-device intelligence and influencing global infrastructure strategies. Despite nine pending lawsuits over copyright issues, Perplexity advances enterprise offerings with numerous integrations and compliance features. The hybrid inference feature, set to launch soon, represents a pivotal evolution in AI orchestration, competing with tech giants deploying similar local-cloud architectures but lacking Perplexity's dynamic task routing capability.

Anthropic's latest browser-based AI model, Opus 4.8, showed a high prompt injection vulnerability, with attackers successfully hijacking it 31.5% of the time before safety measures activated. This figure, detailed in their extensive 244-page system card published on May 28, highlights the challenge of securing AI agents in diverse environments. Unlike OpenAI, Google, and Meta, Anthropic provided transparent, per-surface data across four operational surfaces, including browsers, coding, and tool use, revealing a stark contrast in exposure rates and protective effectiveness. For instance, in coding environments, the attack success dropped from 7.03% to 2.09% once safeguards engaged. Meanwhile, OpenAI reported a robustness score without direct attack rates, Google provided qualitative resistance claims without numeric data, and Meta focused on grading guardrails rather than direct model vulnerability. The article urges security teams to diligently assess AI agent deployment surfaces, request detailed attack success metrics from vendors, and implement rigorous red-teaming themselves prior to deployment to mitigate risk. It underscores that no unified industry standard exists for measuring prompt injection risk yet, making buyer vigilance crucial.

Chinese AI startup MiniMax unveiled its advanced M3 large language model, boasting cutting-edge coding and agentic abilities with a remarkable 1-million-token context window and native multimodality. Offered at subscription prices starting from $20 per month, MiniMax-M3 outperforms leading U.S. proprietary models like GPT-5.5 and Gemini 3.1 Pro on key benchmarks, while costing only 5-10% as much. The company plans to release the model under an open-source license with open weights, enabling enterprises to download, customize, and run it locally free of charge within days. The core innovation, MiniMax Sparse Attention (MSA), drastically reduces computational costs by intelligently filtering attention blocks, resulting in up to 15x faster processing compared to alternatives.

This model excels in autonomous software engineering, scoring 59.0% on SWE-Bench Pro, surpassing many competitors. MiniMax-M3 supports efficient multi-agent coding workflows in its MiniMax Code AI agent, allowing complex tasks to be executed autonomously for extended periods. With flexible subscription tiers catering to different usage levels, MiniMax-M3 offers enterprises full control, enhanced privacy, and cost savings by avoiding cloud API dependency. Early community feedback highlights its strong long-horizon reasoning and highly competitive cost-performance ratio. MiniMax’s open-weights approach reshapes the balance of power in enterprise AI, making frontier-level capabilities accessible with full local control and no vendor lock-in.

In 2024, University of Illinois researchers demonstrated that GPT-4 could autonomously exploit 87% of common vulnerabilities with detailed CVE descriptions, but only 7% without. This offered a safety margin since AI couldn't discover new vulnerabilities—until recently. On April 7, Anthropic's Claude Mythos Preview crossed that margin by autonomously discovering thousands of zero-day vulnerabilities in major systems, rapidly shrinking exploitation timelines. Attacks on vulnerabilities like Langflow and Marimo occurred within hours of disclosure, far outpacing traditional patch cycles. Current defensive infrastructures, largely reliant on slower patch prioritization methods like CVSS, are inadequate. A recently validated three-layer filter combining CISA KEV listings, EPSS scores, and CVSS ratings offers a breakthrough approach to prioritize patches for maximum impact and efficiency. Additionally, organizations must address agent authorization gaps revealed by vulnerabilities like Docker's silent plugin bypass and map the credential blast radius of AI tools to mitigate risks. Practical steps recommended include automating the three-layer filter, deploying event-driven patching for critical services within hours, testing authorization boundaries at scale, documenting credential exposures, and scanning for unauthorized AI agent activity. The rapid evolution of AI-powered exploits demands urgent changes in enterprise security practices beyond standards that take years to materialize. Implementing these measures now can dramatically reduce exposure in this new high-speed threat landscape. Nik Kale, principal engineer in enterprise AI security, underscores the pressing need for transformation in patch management.

Enterprise AI agents are encountering significant delays not due to the performance of AI models but because of restrictive permissioning systems. These agents frequently hit a roadblock when determining what actions they are authorized to perform, on whose behalf, and how authorization is verified. Workday addresses this issue by using its existing system of record as the governance layer for AI agents. Gerrit Kazmaier, president of product and technology at Workday, emphasized that customers struggle when piecing together AI solutions without proper integration, leading to diminished security and overly broad results.

Workday’s solution, Sana, launched in partnership with Google, integrates deeply with enterprise systems to ensure proper authorization and control, especially critical for HR and finance operations where even minor errors can have serious consequences. The system leverages organizational context, identity verification, and layered business process logic to maintain accuracy and compliance. By authenticating users and restricting agent actions within their permitted scope, Workday enhances security and accountability. Industry experts stress that embedding permissions directly within the systems of record is essential to avoid chaos and ensure governance in regulated spaces.

As enterprise AI agents transition into production, organizations are facing significant reliability challenges. Success depends not just on large language model (LLM) performance, but on ensuring long-running AI workflows can survive interruptions, maintain state, recover from failures, and manage costs while coordinating across APIs and tools.

After an initial period of rapid deployment, companies are now revisiting early agent designs to incorporate workflow orchestration, governance, observability, and recovery mechanisms, according to Preeti Somal, Senior VP Engineering at Temporal Technologies. Many are rebuilding their AI systems on more robust foundations to avoid crashes and costly restarts.

This shift highlights the complexity of agentic AI, which often involves multi-step processes spanning various models and services. Handling state—which tracks workflow progress—and memory—which captures ongoing context—is crucial for long-duration business processes, such as healthcare workflows where multiple stages of data processing occur.

Temporal’s approach centers on a "deterministic spine" that manages reliable workflow execution even when AI outputs vary. This helps enterprises avoid failures that would otherwise force expensive, time-consuming reruns. Additionally, orchestration provides clear visibility into where AI-related expenses accumulate, helping companies control inference costs.

Governance and internal frameworks that offer guardrails without stifling flexibility are also increasingly important as enterprises mature in their use of agentic AI. Temporal's platform, already part of many enterprises’ infrastructure, naturally extends to support the evolving needs of AI-driven workflows.

Summary:
Enterprises deploying AI agents in production face complex reliability and cost challenges that require robust architecture beyond the AI models themselves. Solutions focusing on workflow orchestration, failure recovery, state management, and cost visibility are essential for long-running agentic AI processes. Leaders are rebuilding earlier implementations, emphasizing durable systems that can recover from interruptions and provide governance, driving more effective and efficient AI operations.