Anaconda CEO Says AI Agents Are Shipping Faster Than Anyone Can Approve Them

Ask a CIO what’s holding back an AI deployment, and almost nobody says the models aren’t smart enough. Instead, they point to legal teams that haven’t signed off, CISOs unwilling to approve agents with write access, or board members asking what happens when the system gets it wrong—and no one able to give a satisfactory answer. After two years of racing to build smarter models, the AI industry has solved for capability. The harder problem now is accountability.
Gartner expects more than 40% of agentic AI projects to be scrapped by the end of 2027 because of rising costs, unclear business value and inadequate risk controls. Palantir recently warned CIOs that whoever hosts an AI system could ultimately control an organization’s institutional knowledge.
Anaconda CEO David DeSanto is betting on that shift with a string of acquisitions.
Since late April, the Python software company—whose distribution is installed at 95% of Fortune 500 companies—has acquired three startups. The first was Outerbounds, the company behind Netflix’s Metaflow orchestration framework. Next came Kilo Code, an open-source agentic coding platform that grew to 3 million developers in just 16 months. On Aug. 4, Anaconda acquired Enkrypt AI, a Boston startup founded by two Yale Ph.D.s that develops AI red-teaming and guardrail technology.
DeSanto shared that the Enkrypt acquisition was valued in the mid-eight figures and that Anaconda is now worth more than $2 billion.
“Anaconda has been, and continues to be, the trusted foundation for enterprise AI. This has been limited to the development side of enterprise AI and not extending into the production side,” DeSanto told me. “At the beginning of 2026, we set a larger strategy to support the entire AI-native development lifecycle. We determined that strategic acquisitions would help us achieve our goals earlier as enterprises are struggling today and need a trusted partner.”
Enterprise AI’s Next Battle Is Governance, Not Better Models
Enterprise AI still has its familiar scorecard—benchmark rankings, context-window sizes and the cost of processing a million tokens. Those numbers keep improving, and they say less every quarter about whether an organization can actually put AI into production. Spend scales with usage, usage scales with agents, and agents scale with how many people can spin one up—which, thanks to no-code builders across every major platform, is now everyone. Meanwhile the enterprise’s ability to see what those agents touched, what data left the building, and which policy applied at 3 a.m. scales with nothing at all.
DeSanto frames the three acquisitions as three coverage holes in one lifecycle. “Outerbounds provides the production side of AI-native development, Kilo provides agents across the lifecycle, and Enkrypt provides red-teaming and agent guardrails to ensure AI agents do not introduce risk or safety concerns. All of this is built on top of Anaconda’s trusted foundation.”
Three acquisitions in roughly 100 days leave Anaconda with four codebases, roadmaps and support organizations. I asked DeSanto what prevents customers from ending up with three promising products that still require stitching together. “Our original platform and Outerbounds’ former platform are almost done being integrated together into a single platform and it is available today in beta to organizations who would like to adopt it before our GA at the end of September,” he says. He also disclosed that Kilo Code’s integration with Anaconda Desktop has already shipped. The Enkrypt acquisition, meanwhile, may prove the most consequential despite being the smallest of the three. The deal adds AI red-teaming across more than 300 attack categories, runtime guardrails against jailbreaks and data leakage, and compliance mapping for the NIST AI Risk Management Framework and the EU AI Act.Â
Enkrypt is an OpenAI compliance integration partner, allowing ChatGPT Enterprise customers to connect its audit and guardrail controls directly into their deployments, with Anthropic support planned. The platform claims it has scanned more than 268,000 AI tools across 25,000 Model Context Protocol servers over two months, identifying more than 143,000 vulnerabilities affecting 73% of them. The figures were published the day the acquisition closed, but without an accompanying methodology. So, what exactly counts as a vulnerability? DeSanto says the company scanned publicly available MCP servers and exposed tool definitions using an automated taxonomy before manually validating a sample. “Our focus is AI-specific vulnerabilities, not generic infrastructure flaws,” he says.
Path traversal accounted for 31.7% of findings, while resource exhaustion represented 24.5%. Together, those two categories made up more than half of all reported vulnerabilities, yet both are well-known infrastructure weaknesses that happen to surface through MCP tool interfaces. Prompt injection represented 21% of findings.
“We don’t treat all 143,000 as equally severe,” DeSanto says. “Path traversal and resource exhaustion are common but often lower-severity hygiene issues, while prompt injection and command injection are what enterprises should weigh most heavily.” Enterprises, he notes, should ask every security vendor—including Anaconda—for a category-by-category breakdown rather than a single headline number.
The platform now cites 52 million users, up from 50 million a year ago. DeSanto outlined how the company arrives at that figure. “Anaconda serves more than 52 million users who are using the conda package manager, our desktop products, Anaconda Core, or have downloaded package details for items like software supply chain reporting,” DeSanto says. “We are working to include Kilo users, which now is more than 3 million, and will include Enkrypt users in the coming months. Altogether, Anaconda serves more than 55 million users worldwide.”
Reach, not revenue, and the jump to 55 million comes by addition rather than adoption.
The Next AI Security Challenge Is Agents That Build Other Agents
Every control-layer product shipping today assumes an agent executes tasks a human defined. The next generation writes its own tools, spawns sub-agents and rewrites the scaffolding it runs on. That raises a question about whether today’s runtime guardrails and compliance frameworks are built for the next generation of AI agents or the last.
“This means taking existing security models and adding in safety checks, guardrails, and visibility needed to address a fundamentally different paradigm that AI agents introduce,” DeSanto says. “This includes not only ensuring they operate securely, but at enterprise scale without anything breaking or operating in the dark.”
Describing what makes prompt injection the category most specific to agents, DeSanto points to “tool descriptions and inputs crafted to manipulate how an agent reasons or acts, not how a server responds.” That distinction is the whole problem in miniature. An agent that generates its own tool descriptions is generating its own attack surface faster than any taxonomy can enumerate it.
Control Layer Verdict
The enterprise AI race stopped being about which model is smartest somewhere around the second abandoned pilot. What’s left to win is the approval layer—the thing that decides what every model underneath it is permitted to do. Anaconda has bought a real position there in about 100 days, faster than any company can integrate anything. Buyers should know what they’re actually shopping for. Not a smarter system—a shorter list of people who can say no.
Victor Dey is a tech analyst and writer who covers AI, data science, startups, and cybersecurity. A former AI editor at VentureBeat, his work also appears in New York Observer, Fast Company, Entrepreneur Magazine, HackerNoon, and more. Victor has mentored student founders at accelerator programs at leading universities including the University of Oxford and the University of Southern California, and holds a Master's degree in data science and analytics.
More from Victor Dey
See all
Future of AICysic CEO Says Paying for an AI Model Doesn’t Prove Your Host Ran It
Future of AIInside Indeed’s AI Hunt For Fake Employers Before Anyone Applies
Future of AIDatabricks Co-Founder Explains How Genie Tackles the Blind Spot In Always-On AI
Enterprise AIWhat Happens to Enterprise Software When AI Does the Work?
Recommended for you
More stories
AI & TechnologyFrom self-driving excavators to AI leasing agents, proptech’s AI moment has arrived
AI & TechnologyAI literacy shouldn’t begin in the workplace. It should begin in primary school.
DataThe Data Centre Is Becoming a Product
AI & TechnologyToo often, AI narrows rather than widens access. Here’s how the tech industry can fix it
AI & TechnologyHow to Review an AI Subscription Before It Renews
AI & Technology