Autonomous security: who’s watching your AI agents?
Remember when using AI meant typing a question into a chat box and getting a paragraph back? That feels like a long time ago now.
Today, tools like Claude Code, Cursor and Codex don’t stop at answering. They open files, run commands, connect to other apps and finish tasks on their own. Developers use them to ship code. Marketers use them to pull reports. Whole teams lean on them every day.
That is a big upgrade in what we can get done. It also raises a question most of us haven’t stopped to ask. When an AI agent works on your laptop with your logins and your access, who is keeping an eye on what it does?
That question is the reason Autonomous Security exists. The New York startup, which you’ll find at a16y.ai, builds a platform that secures AI agents across endpoints, browsers, SaaS and cloud, with real-time policy enforcement. Its tagline sums up the idea well: Move Fast with AI. Move Faster with Security.

The risk starts after you hit Enter
Here’s the part that changed how I think about AI tools. Your prompt tells the agent what you want. After that, the agent makes its own calls. It decides what to retrieve, where to connect, which credentials to use and what action to take next.
Every one of those steps can look perfectly normal to a security system. The agent is using files, tools and accounts it already has permission to touch. Traditional endpoint security can log a command, but it has no idea why the agent ran it.
Autonomous Security calls this the problem of risk beginning after the prompt. And the company has numbers to back it up. Over one month, it analyzed more than 100,000 enterprise AI sessions. It found agents searching credential stores without being asked, following injected instructions and taking potentially high-impact actions.
None of that requires a hacker sitting in a dark room. A poisoned web page, a sketchy plugin or a badly scoped permission can be enough to send a helpful agent off-script. The company’s full findings are in its report, Why AI Security Fails After the Prompt, which is free to download from its site.
See every agent, then control what it does
Autonomous doesn’t rely on a single scanner or a single gateway. It stacks several controls that work together. The difference is context: the platform ties each action an agent takes back to the prompt and the session behind it. That lets a security team spot when an agent goes off-script and stop an unsafe action before it reaches company systems.
The platform breaks down into four main jobs.
1. Discover Shadow AI. “Shadow AI” is the AI nobody officially signed off on. Think of the coding assistant someone installed on a Friday afternoon, or the browser extension that quietly connects to a chatbot. Autonomous builds an inventory of every AI agent, app, model, account and component in use. Admins can then mark each one as approved, restricted or prohibited.
2. Scan MCP servers and skills. MCP servers and skills are the add-ons that give agents new abilities, like reading your calendar or querying a database. They’re also an easy way for trouble to sneak in. Autonomous continuously checks them for vulnerabilities, malicious behavior, risky permissions and exposed secrets.
3. Guard agents in real time. This is the runtime layer. It watches prompts, tool calls, commands, data access and actions as they happen. Autonomous describes it as fast AI-powered detection paired with deterministic, real-time enforcement. Its built-in DLP (data loss prevention) can automatically restrict an agent’s workflow the moment it touches sensitive data.
4. Keep credentials out of agents. I like this one most. A token vault stores credentials and gives agents access without ever handing over the underlying secrets. Managed identities and secure gateways also govern how agents connect to tools, models, APIs and MCP servers. So even a misbehaving agent has fewer keys to lose.
It goes wherever your agents go
Agents don’t live in one place, so a tool that only watches one place will miss a lot. Autonomous covers six environments under one platform:
- Endpoints: macOS, Windows and Linux machines
- Browsers: Chrome, Edge and Firefox
- AI SaaS: Claude, ChatGPT and Microsoft 365 Copilot
- MCP servers: through the company’s own MCP Gateway
- LLM gateways: guardrails on your LiteLLM setup
- Containers and cloud: Docker, Kubernetes and cloud agents
The company says it covers more than 60 AI agents. The list on its site reads like a roll call of today’s popular tools: Claude Code, Cursor, OpenAI Codex, ChatGPT, GitHub Copilot CLI, Gemini CLI, Windsurf, JetBrains, Grok, Antigravity, Cline, Ollama, Perplexity Comet and plenty more.
Setup is where Autonomous makes a bold promise. It says integration takes about 10 minutes. You install the endpoint agent by hand or push it out through your MDM, and its first scan inventories a machine’s AI agents within a few minutes. Findings then flow into the tools a security team already uses, including a management console, APIs and webhooks, SIEM and SOAR platforms like Splunk and QRadar, OpenTelemetry, Syslog and Slack alerts. Nobody has to learn a whole new dashboard just to get value on day one.
The team has done this before
Autonomous Security didn’t start under that name. It launched in early 2025 as MCPTotal, a company built to make sense of the sudden explosion of MCP servers. Then, in the second half of 2025, AI agents spread through workplaces much faster than anyone expected. The founders saw that almost no security tooling had been built for them.
So they moved to where agents run, starting with the endpoint. That’s the spot where prompts can be seen, tool calls intercepted and secrets vaulted before something goes wrong. With that shift came the new name. The team and the mission stayed the same.
The two people behind it have built a company together once already:
- Gil Dabah, co-founder and CEO, co-founded NorthBit, which Magic Leap acquired in 2016. He’s a security researcher credited with dozens of CVEs (publicly logged software vulnerabilities).
- Ariel Shiftan, PhD, co-founder and CTO, also co-founded NorthBit and won first place in the XPRIZE AI competition in 2021.
Both served in an elite IDF intelligence and cybersecurity unit. Autonomous is their second company together.
The backing is just as notable. Autonomous has raised $15M in total from YL Ventures and George Kurtz, the co-founder and CEO of CrowdStrike. The founders of Snyk, Armis and Wiz have also invested. When people who built some of the biggest names in cybersecurity put their own money in, it says something about where they think the next problem lies.
What security teams are saying
Autonomous already counts SolarEdge, Oleria, Xactly, BigID, The5ers, Crazy Ant Labs and ComeryX among its customers. A few of them have shared how it’s going.
Shlomi Oberman, CISO at SolarEdge, put the scale in plain terms: “Thanks to Autonomous we now have full visibility into Shadow AI and our software supply chain on 4,000 endpoints.”
Matthew K. Sharp, CISO at Xactly, pointed to something easy to overlook. Rolling out AI tools is one thing. Knowing how they’re used months later is another. He said Autonomous gave his team “continuous, fleet-wide visibility into how Claude, Cursor, and Codex are actually being used,” and helped them find security exposures long after the initial deployment.
The quote that stuck with me most came from a smaller team. Moty Michaely, CTO and founder of Crazy Ant Labs, said: “With a fully remote team doing 100% of our development with AI, we needed to see and control that AI. Autonomous gave us both, and we ship faster and safer.”
That last line matters. Security tools often get a reputation for slowing people down. Here, the pitch is the opposite.
Who should look into it, and how to start
Let’s be clear: Autonomous is built for organizations. It isn’t an app you install on your personal laptop to babysit a chatbot. Still, a lot of Gadget Flow readers fit the profile. If you run a startup, lead a dev team, manage IT for a growing company or work in a team where everyone has quietly started using AI agents, this one is worth a look. It’s also a smart read for anyone curious about where AI is heading, because agent security is going to be a much bigger conversation in the next few years.
There are a few easy ways to get started:
- Free MCP and skills scan: You can scan your MCP servers and skills at no cost through the Autonomous web app.
- 7-day free trial: Full access, and no credit card needed.
- 30-minute demo: The demo runs with the founders on a real environment, and the site makes a point of saying it’s “not a sales team.”
Pricing isn’t listed publicly. The base subscription is fixed and tiered by the number of endpoints you monitor, with à la carte options on top, so you request a quote. Enterprise buyers will also be glad to see SOC 2 Type II compliance, SSO and SCIM, role-based access, audit logs and a guaranteed support SLA.
The bottom line
AI agents are some of the most useful tools we’ve picked up in years. They’re also running with more access than most of us realize. Autonomous Security gives teams a way to see every agent, understand what it’s doing and step in before a helpful assistant turns into an expensive mistake.
If your team is going all in on AI, this is the kind of safety net to set up early. Check out Autonomous Security at a16y.ai to run a free scan or book a demo.




