Amazon Web Services (AWS) has announced a significant upgrade to its AI-driven security platform, Continuum, integrating it directly with OpenAI's Codex and Anthropic's Claude Code. This integration embeds AWS security tools within the coding environments of these AI rivals, creating a unified security control plane for enterprise software development in the AI era. Continuum uses advanced AI models to identify, prioritize, validate, and remediate code vulnerabilities automatically, optimizing the use of various models for different tasks while absorbing the token costs. Additionally, AWS has expanded its Security Hub Extended by adding a new supply chain security category with partners Chainguard and Socket, offering a comprehensive solution against modern open source threats. This dual approach highlights AWS's strategic focus on owning the orchestration layer that connects AI models, open source packages, and security solutions, aiming to make security a seamless part of the developer workflow while addressing emerging threats such as unregistered AI agents and cost harvesting attacks. AWS's ambition is to become the indispensable platform securing the AI software supply chain across industries, emphasizing autonomous security operating at machine speed.
Brex CEO Pedro Franceschi presented at VB Transform 2026 an innovative approach to securely deploying AI agents like OpenClaw within enterprises. Viewing AI not as vague agents but as virtual employees, Brex created a security model shifting focus from controlling agent code to monitoring network traffic. Their open-source proxy, CrabTrap, watches outbound network connections and uses a large language model (LLM) to judge if actions comply with policy, flagging suspicious behavior for human review. This approach acknowledges AI agents' coding power while ensuring security through network-level oversight, offering a new paradigm for safely integrating AI into corporate environments.
Meta has introduced Muse Glimmer, a 30-billion-parameter AI model designed to run autonomous agents directly on high-end consumer hardware like Macs and PCs, reducing reliance on cloud infrastructure. This model marks Meta’s return to open-source with a fully Apache 2.0 licensed release—offering unrestricted commercial use, modification, and redistribution, unlike the previous Llama model with more restrictions. Muse Glimmer supports multimodal inputs (text and images), operates with over 100 languages, and is optimized for local deployments on 24–32GB VRAM systems through quantized versions and advanced decoding techniques. It’s built for agentic tasks—planning, tool use, interpretation, and failure recovery—and shows competitive performance compared to other large models like Google’s Gemma 4 and Alibaba’s Qwen, specializing in reliable local AI agents. Meta is actively collaborating with hardware partners and has published detailed developer documentation to support the ecosystem. The open-source release promotes safer, local AI use coupled with performance efficiency, pushing forward the feasibility of complex AI agents working entirely on-device.
Content filters are designed to block unsafe outputs but cannot determine if an AI agent was authorized to perform actions like issuing refunds or making changes in production systems. This gap poses a distinct challenge that many enterprises have yet to address adequately. AI agents may follow their instructions precisely yet take actions beyond the business’s sanctioned authority, leading to operational risks and compliance issues.
In practical business scenarios, agents might calculate refunds correctly yet exceed approved limits or apply changes without considering financing or fulfillment constraints. These are not errors in AI reasoning but failures to clearly separate technical capabilities from business decision rights.
As AI moves from recommending assistants to agents that execute workflows, each production agent must have explicit decision rights outlining what it can execute, recommend, or must avoid. Guardrails that restrict behavior don’t equate to authority models.
Safety controls manage harmful content and behavior but don’t address whether an agent is authorized to act for the enterprise. A 2026 Cloud Security Alliance survey revealed a significant governance gap, with many enterprises unaware of AI agents operating autonomously in their environments.
An effective approach involves creating an Agent Authority Contract — a machine-enforceable record detailing the scope of an agent’s delegated power, including ownership of outcomes, permitted actions, system access, materiality limits, escalation triggers, reversibility, and duration of authority.
Every significant agent action should be categorized into one of four outcomes: Allow (low-risk autonomous actions), Approve (actions requiring human or policy approval), Recommend (agent proposes actions for human decision), or Deny (actions outside the agent’s authority). These decisions must be enforced beyond just system instructions.
Authority decisions should be dynamic, made in real-time based on context, identity, and potential impact. Human oversight should focus on exceptions and high-risk cases rather than all actions, balancing autonomy with responsible governance.
Measuring metrics like override rates, escalation accuracy, unauthorized attempts, and business-impact errors can help enterprises maintain an optimal balance of AI authority and control.
The key governance challenge is not AI models or safety controls but defining who delegates authority, the scope of delegation, and how it is enforced and observed. Enterprises must clearly determine what they are willing to delegate before adopting autonomous agents fully.
AI's threat in medicine is not just about experts losing their reasoning skills, but about medical students and trainees who may never develop these critical skills due to early dependence on AI tools. While practicing doctors might become less adept clinically with AI, the greater danger lies with trainees who use AI before building their own judgment. Unlike deskilling, which implies loss of an ability, this scenario risks 'never-skilling'—meaning some doctors might not learn how to reason clinically at all. Tools like OpenEvidence, an AI chatbot used by two-thirds of U.S. doctors for quick, research-based answers, are now also popular among medical trainees. This early reliance could hinder the essential development of independent clinical reasoning during formative education stages.
Amazon is planning a new data center in Texas, which will include an on-site power plant. This facility has the potential to become the largest single source of climate pollution in the United States, raising significant environmental concerns.
NextSlide has announced that its team members have joined OpenAI and are currently contributing to the development of ChatGPT.
In July, startup investors maintained strong momentum, with well-known names dominating in both deal numbers and investment sizes. Khosla Ventures led in active lead investments with eight deals over $5 million, including major rounds for Oratomic and Norm AI. Y Combinator stood out as the most active overall investor, participating in 19 deals, mainly as a non-lead investor. Coatue made the largest expenditures, notably backing Blue Origin’s $10 billion financing, while Nvidia invested $5 billion in Safe Superintelligence. Other key players included Insight Partners, Andreessen Horowitz, and Index Ventures. Seed stage investments were led by Y Combinator, alongside LvlUp Ventures and Alumni Ventures.
Following significant AI investment, Rippling has introduced the AI Spend Console, a new tool designed to track AI-related expenditures at both individual and team levels, helping businesses assess the return on investment in AI resources among employees.
Kitesurf is a cloud-based browser built specifically for AI agents rather than human users. It operates with lower computational demands compared to Chromium for typical automation tasks, enabling developers to create browser-driven AI agents more efficiently.
