Anaconda Joins Open Secure AI Alliance to Tackle Agent and MCP Risks

Why Anaconda points to MCP server risk

According to Anaconda, AI agents act on business data and enterprise systems through the tools they call, and those tools can carry security risk. The company says that in two months, Enkrypt AI by Anaconda scanned 268,000+ tools across 25,000 Model Context Protocol (MCP) servers. The scan uncovered 143,000+ vulnerabilities affecting 73% of those servers.

Anaconda frames the issue as a practical one for enterprise leaders who want their teams to do more with agents: how to scale the value while keeping control. The company says it has secured and governed the open-source software supply chain for enterprise data science, machine learning and AI for more than a decade, and it brings that experience to the alliance.

What the alliance does

The Open Secure AI Alliance brings together technology companies, researchers and government partners to develop and share open tools, standards and techniques for securing AI. Anaconda lists NVIDIA, Microsoft and AWS, all Anaconda partners, among the members, alongside Cloudflare, Red Hat and others across enterprise technology and cybersecurity.

The alliance's approach centers on giving organizations defensive capabilities they can inspect, adapt and run on infrastructure they control. Its scope covers the full agent stack: models, the software that lets agents take action, and the policies and controls that govern them.

Anaconda argues that security must reach beyond the model itself. Teams also need to understand the packages a system depends on, the tools it can reach, the data it can access and the actions it can take. As an example, the company describes an agent that retrieves a document, changes its plan based on what it reads, and then updates a business record. Each step may look reasonable alone, but the full sequence can still take the agent outside the scope the business approved.

Three priorities and the Anaconda Platform

Anaconda says it plans to work with other alliance members on three priorities:

• Test agents across their full chain of decisions. Improve evaluation methods and red-teaming techniques to find failures that unfold across multiple steps.
• Secure the tools and MCP servers agents connect to. Advance common approaches to scanning, vetting and controlling access across the AI software supply chain.
• Turn AI policy into evidence. Improve how organizations record agent decisions and report incidents, so security and compliance teams can verify which controls ran and what they allowed or stopped.

The company also ties the move to its Anaconda Platform, which it describes as four capabilities for building, securing and operating AI. AI workspaces use Kilo by Anaconda, with model choice, trusted packages and models, and visibility into usage and cost. Kilo Enterprise adds model and provider controls, single sign-on (SSO) and audit logs. AI security and guardrails rely on Enkrypt AI by Anaconda to red-team agents before release, enforce guardrails on agent tool calls in production and keep audit-ready evidence. AI orchestration provides reproducible, traceable workflows across environments. AI artifacts offer vetted Python packages, curated models and trusted MCP servers, with provenance and vulnerability tracking.

What it means for enterprise teams

Anaconda says it supports open standards for testing agents, vetting MCP servers and recording agent decisions, and plans to share what it learns with the wider community. Its stated goal is for organizations to scale AI on their own terms, with builder freedom and enterprise control, using open tools and standards the community can inspect and build on.

For enterprises that mix open and proprietary models, development tools, data platforms and infrastructure, the company argues that shared security approaches can help them evaluate and protect those systems while keeping the freedom to choose technology. The source gives no timeline or specific deliverables for the alliance's work, so the practical outcomes remain to be seen.