Today on The Gateway Signal, enterprise AI adoption has quietly built a massive supply chain problem. A new Snyk report reveals the average corporate AI footprint is now three times larger than its actual model inventory, thanks to an expanding web of agents, frameworks, and databases. We're tracking how the industry is scrambling to secure this broader attack surface, starting with new initiatives from Chainguard and Databricks.
Databricks announced on Thursday that it has officially integrated the skills from its AI Dev Kit into its CLI-installed 'Databricks AI Tools'. Following the string of critical vulnerabilities we've tracked in LiteLLM, Databricks explicitly noted it is moving away from the open-source gateway for core functionality to secure its supply chain. The AI Dev Kit will now focus entirely on the recently modernized Model Context Protocol (MCP) server.
Why it matters
This move by Databricks signifies the maturation of enterprise AI development, consolidating tools under an officially supported, more secure umbrella. The explicit mention of moving away from LiteLLM for security reasons is a significant signal for the open-source gateway space, highlighting that enterprises are prioritizing supply chain integrity. It also reinforces the growing importance of MCP as a standardized, governed layer for agent-tool interaction.
Building on Y Combinator's recent open-sourcing of its 'QM' agent harness, CEO Garry Tan launched GBrain on Thursday, an open-source 'brain layer' for AI agents. The tool acts as a persistent memory that synthesizes information with citations, performs gap analysis on its own knowledge, and builds a self-wiring knowledge graph of entities and relationships. It can be integrated with coding agents like Claude Code and Codex or run as a full autonomous agent, and it deeply integrates with the Model Context Protocol (MCP).
Why it matters
GBrain addresses a fundamental limitation of many current AI agents: the lack of persistent, structured memory and advanced reasoning. By providing a framework for synthesis and gap analysis, it moves beyond simple Retrieval-Augmented Generation (RAG) to create agents that can build and query their own understanding of a domain. As an open-source project from a high-profile figure, it's likely to see significant adoption and influence the architecture of future agentic systems and developer tools.
The terminal company Warp has released a standalone Agent CLI that gives its AI agent native control over shell operations. A novel pseudoterminal multiplexer architecture allows the agent to interact with long-running, interactive applications such as REPLs and debuggers, a capability that, according to the company, is unmatched by other AI coding agents that typically only handle single commands.
Why it matters
This is a significant architectural innovation in the AI developer tool space. By solving the problem of interacting with stateful, interactive terminal sessions, Warp's agent can automate a much wider and more complex range of developer tasks, from database operations to live debugging. It positions Warp not just as a terminal, but as a potential orchestration layer for multi-model coding agents, moving beyond simple code generation to true workflow automation.
Blackstone is reportedly exploring a debt package of at least $36 billion to finance Anthropic's use of Google's AI chips, according to reports on Wednesday. This is part of a complex, off-balance-sheet financing structure engineered by Google with partners including Broadcom, Apollo, and Morgan Stanley to provide Anthropic with chips and data center capacity, reportedly involving up to $200 billion in contracts sitting in a separate vehicle.
Why it matters
This massive and complex financial engineering highlights the staggering capital expenditure required for frontier AI development. It shows that the AI infrastructure buildout is now so large that it requires creative, syndicated financing from private credit markets, moving beyond traditional venture capital or corporate balance sheets. This trend establishes AI infrastructure as a distinct, highly capitalized asset class.
DeepSeek is reportedly preparing to launch 'DeepSeek Harness,' an extensible foundational architecture for AI agents, aiming to foster an open ecosystem around its high-performance models. The news, from a Wednesday report, comes as a separate report claims the company—whose $52 billion valuation and 2027 IPO plans we tracked last month—has reopened talks for a massive RMB 50 billion (~$6.9B) funding round that was previously suspended.
Why it matters
This signals a major strategic push by DeepSeek to move up the stack from a model provider to a platform player, creating an ecosystem that could compete with OpenAI's GPTs. For the Chinese AI scene, it represents a clear ambition to not just produce cheap, powerful models, but to also define the application layer infrastructure. If the funding round is successful, DeepSeek will be heavily capitalized to pursue this strategy aggressively.
Following its entry into the paid API market last month with Muse Spark 1.1, Meta launched its 'Muse Code' coding agent on Wednesday, powered by a new Muse Spark 1.2 foundation model. Developed by Meta Superintelligence Labs, the tool is positioned to compete with OpenAI's Codex and Anthropic's Claude. It is available on multi-model platforms like OpenRouter and includes a zero-data retention option for enterprise clients.
Why it matters
Meta's entry adds another major player to the increasingly competitive coding agent space, while its zero-data retention policy directly addresses enterprise privacy concerns. Its immediate availability on OpenRouter further validates the multi-model gateways we've been tracking as a primary distribution channel for reaching developers.
Cloudflare on Wednesday launched and open-sourced 'Cloudflare OS,' an AI workspace designed to run on a company's own Cloudflare account. Released under an Apache 2.0 license, the platform allows employees to securely build custom micro-apps and automate workflows. It integrates with the Cloudflare AI Gateway, enabling organizations to use any AI model provider and route tasks based on complexity and cost.
Why it matters
Cloudflare OS is a significant new entry in the open-source AI infrastructure landscape. It provides enterprises a self-hostable, secure alternative to closed, proprietary AI work platforms. The tight integration with its AI Gateway reinforces the gateway's central role in a multi-model strategy, giving platform teams a powerful, open-source tool for managing AI usage, costs, and governance without being locked into a single vendor's ecosystem.
Amid ongoing debate over its competitive moat and a rumored $10 billion acquisition by Stripe, AI gateway OpenRouter has released 'Ori,' a command-line tool that pre-configures four different coding agents (Claude Code, Codex, OpenCode, Hermes Agent) with a single command. The tool sets specific, optimized per-model settings, most notably enabling the 'tool search' feature for Anthropic models to significantly reduce system prompt token counts and improve agent behavior.
Why it matters
This is a clear move by OpenRouter to differentiate itself beyond simple routing and pricing. By providing a developer experience tool that handles model-specific optimizations, it adds a layer of value that makes using its gateway more efficient than calling the models directly. This highlights a trend where gateways are evolving into intelligent platforms that abstract away the complexities of the underlying model ecosystem.
ngrok launched its AI Gateway, 'ngrok.ai,' on Thursday, offering a unified platform for managing requests across public providers like OpenAI and Anthropic, custom endpoints, and self-hosted models. The service provides observability, access control, and fallback logic, specifically enabling developers to route sensitive workloads to self-hosted models via a secure tunnel.
Why it matters
The ngrok AI Gateway directly addresses the hybrid infrastructure needs of many enterprises. Its key differentiator is simplifying the integration of private, self-hosted models with public APIs, a common architectural pattern that can be complex to manage. By leveraging its core secure tunneling competency, ngrok provides a practical solution for companies that want to maintain data privacy and control for specific tasks while still using powerful public models for general-purpose needs.
A new Snyk report published Wednesday reveals that the average enterprise AI footprint is three times larger than their model inventories indicate, due to the adoption of agentic architectures. This expanded ecosystem includes frameworks like the Model Context Protocol (MCP), vector databases, retrieval systems, and other third-party tools, with nearly half of organizations consuming AI through these dependencies rather than direct model APIs.
Why it matters
This report quantifies a critical shift: the challenge of production AI is no longer just managing the model, but governing its entire, sprawling supply chain. For platform and security teams, this means the attack surface is much wider than assumed, and governance must extend beyond API gateways to encompass every framework, database, and tool an agent might touch. This validates the need for dedicated AI security solutions that can provide visibility into these complex dependencies.
Chainguard introduced its Agent Skills initiative on Thursday, a new product aimed at securing the software supply chain for AI coding agents. The initiative provides a public registry of hardened, auditable agent skills and a private registry for internal enterprise use, treating AI agent capabilities as first-class software artifacts that require continuous hardening, governance, and provenance tracking.
Why it matters
As agentic AI moves into production, the skills and tools they use represent a new, significant attack surface. Chainguard's initiative is one of the first major commercial efforts to address this emerging security gap, shifting the focus from securing the model to securing the actions the model can take. This is a critical infrastructure layer for any enterprise deploying autonomous agents, directly competing with and complementing governance features within AI gateways.
On Wednesday, AMD announced its Helios rack-scale AI system has entered full production, with first shipments expected by the end of Q3 2026. Microsoft is confirmed as the newest hyperscaler to deploy Helios on Azure. The system, which competes directly with Nvidia's rack-scale offerings, features 72 Instinct MI455X GPUs and is also being adopted by Meta, OpenAI, and Anthropic.
Why it matters
Microsoft's adoption of Helios is a major validation for AMD and a significant development in the AI hardware market. It provides a credible, large-scale alternative to Nvidia's dominant platform, giving hyperscalers and major AI labs a second source for their massive compute needs. This increased competition is critical for the long-term health of the AI infrastructure market, potentially leading to better pricing and supply chain resilience.
Enterprise AI Footprint Extends Far Beyond Models A new Snyk report finds that the true AI footprint in enterprises is 3x larger than model inventories suggest, encompassing a wide array of frameworks, databases, and tools, many from third-party services. This hidden complexity is driving the need for comprehensive governance and supply chain security, as evidenced by new security offerings from Chainguard and the consolidation of tools by Databricks.
AI Agent Infrastructure and Security Emerge as Key Investment Areas The focus is shifting to building and securing the infrastructure that surrounds AI agents. Chainguard has launched a registry for hardened agent skills, while Databricks is integrating its own AI tools. Concurrently, Y Combinator's Garry Tan released GBrain, a 'brain layer' for agents, and DeepSeek is reportedly developing its own agent framework, 'DeepSeek Harness,' indicating a push towards more capable and secure agent ecosystems.
The Chinese AI Ecosystem Accelerates with Open Models and Agent Frameworks Chinese AI labs are rapidly innovating and competing on the global stage. DeepSeek is reportedly developing an open agent framework called 'DeepSeek Harness' and is rumored to be reopening a massive funding round. This comes as analysis from 36Kr highlights how models like Moonshot's Kimi K3 and DeepSeek-V4-Flash are dominating global rankings in usage and performance, driving down prices worldwide.
Open-Source AI Tooling Matures with New Releases and Enterprise Adoption The open-source AI ecosystem continues to produce robust tools for developers and enterprises. Garry Tan of YC has released 'GBrain,' an advanced knowledge layer for agents. AWS's Kiro Crew agent orchestrator is now open-sourced, and Cloudflare has released 'Cloudflare OS,' an open-source AI workspace. These tools provide powerful, self-hostable alternatives to commercial platforms.
Financing for AI Infrastructure Reaches Unprecedented Scale The capital required to power the AI boom is leading to massive and complex financial deals. Blackstone is reportedly exploring a $36 billion debt package for Anthropic's chip needs, part of a larger trend where tech giants and private credit firms create off-balance-sheet vehicles to fund AI compute. This underscores the immense, ongoing investment required to build out foundational AI infrastructure.
What to Expect
Late August 2026—DeepSeek's potential RMB 50 billion second funding round could be finalized.
How We Built This Briefing
Every story, researched.
Every story verified across multiple sources before publication.
🔍
Scanned
Across multiple search engines and news databases
446
📖
Read in full
Every article opened, read, and evaluated
169
⭐
Published today
Ranked by importance and verified across sources
12
— The Gateway Signal
🎙 Listen as a podcast
Subscribe in your favorite podcast app to get each new briefing delivered automatically as audio.
Apple Podcasts
Library tab → ••• menu → Follow a Show by URL → paste