Builders get a single platform for running coordinated swarms, vetting open-weight models, and red-teaming those agents across more than 300 attack categories before code reaches production. The expansion of the Anaconda Platform links agentic development with autonomous security testing, so the same environment that coordinates multiple coding helpers is now the one that probes them for weaknesses. The rollout is built on three capabilities acquired over recent months: Kilo Code, Enkrypt AI, and Outerbounds.
Why the timing matters
Anaconda’s own survey of AI-native builders found 63% of respondents moving toward agent swarms in some form. That is a clear signal that swarms are shifting from experiment to default workflow, which raises the cost of shipping insecure code. As agents gain access to enterprise tools and data, the surface that needs testing grows with it.
That gap shows up in independent research. A four-month scan of more than 268,210 agent tools across 25,264 MCP servers found vulnerabilities in 73% of them. Builders need to see which agent ran, what data it touched, which tool it called, and whether it stayed inside its assigned authority.
What is new on the platform
- AI Workspaces: a shared environment that brings together agentic tools and vetted components for builders and data scientists, with model choice and cost controls built in.
- Agent Swarms in VS Code: through Kilo, the primary AI workspace, multiple agents now coordinate in parallel, share context, and optimize token use. Local development is handled by Kilo Desktop, which combines software engineering, data science, and secure Python environment management with access to 500+ AI models, 19,000+ vetted packages, local model execution, and live notebook sessions where users and agents edit and run cells together. Sign in with ChatGPT gives access to every Kilo surface without extra logins or added token costs.
- Source-built packages: 13,000+ newly vetted AI, ML, and Python packages added to the existing source-built library.
- Model Catalog: 77 vetted open-weight models in a curated library, expanding builder choice without giving up governance.
- MCP: the same trust and governance extended to agent tool calls through the new Anaconda MCP.
- AI Security and Guardrails: continuous security testing and runtime protection. Red-teaming challenges models, agents, and MCPs across 300+ attack categories and adapts in real time to surface weaknesses in deployment and production. Guardrails approve, modify, or block risky behavior across agents, tools, RAG, and MCP.
- Agent Incident Registry: a source-backed record of publicly reported agent incidents, offering enterprises an independent way to confirm security claims beyond vendor statements.
- AI Orchestration: repeatable workflows and reproducible environments that move AI from development through testing into production. Interactive inference deploys models straight from the catalog with no separate step, and FastBakery builds reproducible container images that compile conda and PyPI dependencies, including native libraries, into images pip-only tools cannot produce.
Why security has to move at the same speed as the agents
Independent analyst research found that 72% of organizations rank managing growing agent autonomy as critical or very important to their AI strategy. The pattern is consistent: enterprises that scale it successfully separate themselves from those that stall by keeping security and visibility at the same pace as the agents themselves. Continuous testing, runtime guardrails, and an independent incident record together give builders a way to verify what their systems are actually doing in production.
What this means in practice
For a builder, the change is concrete. Swarms can coordinate several workstreams at once instead of one after another, which turns a single day’s output into a small team’s output. The trade-off is that every coordinated agent is a new attack surface, so red-teaming and guardrails have to run alongside the work rather than at the end of it. The Agent Incident Registry adds a public record that lets teams confirm security claims outside of vendor messaging.
For an enterprise, the practical gain is a single environment that holds the packages, the models, the orchestration, the red-team coverage, and the runtime controls, so governance travels with the workflow from development through production.
FAQ
What is the Anaconda Platform expansion announced in October 2026?
The expansion pairs agentic development with autonomous security testing, bringing agent swarms, autonomous red-team agents, vetted open-weight models, and runtime guardrails together on one platform.
How are agent swarms delivered on the platform?
Agent swarms reach VS Code through Kilo, the primary AI workspace. Multiple agents build together in parallel, share context, and optimize token costs. Kilo Desktop adds local model execution, access to 500+ AI models and 19,000+ vetted packages, and live notebook sessions where users and agents edit and run cells together.
What does the new autonomous security testing cover?
Red-teaming challenges models, agents, and MCPs across more than 300 attack categories and adapts in real time, while guardrails approve, modify, or block risky behavior across agents, tools, RAG, and MCP at runtime. The Agent Incident Registry provides a source-backed record of publicly reported agent incidents for independent verification.
This article summarizes reporting from helpnetsecurity.com.