LLM Entry With out the Trouble
DevNet Studying Labs give builders preconfigured, in-browser environments for hands-on studying—no setup, no surroundings points. Begin a lab, and also you’re coding in seconds.
Now we’re including LLM entry to that have. Cisco merchandise are more and more AI-powered, and learners have to work with LLMs hands-on—not simply examine them. However we are able to’t simply hand out API keys. Keys get leaked, shared outdoors the lab, or blow via budgets. We wanted a solution to prolong that very same frictionless expertise to AI—give learners actual LLM entry with out the chance.
At this time, we’re launching managed LLM entry for Studying Labs—enabling hands-on expertise with the newest Cisco AI merchandise and accelerating studying and adoption of AI applied sciences.
Begin a Lab, Get On the spot LLM Entry
The expertise for learners is straightforward: begin an LLM-enabled lab, and the surroundings is prepared. No API keys to handle, no configuration, and no signup with exterior suppliers. The platform handles all the things behind the scenes.
The quickest path at this time is A2A Protocol Safety. Within the setup module, the lab hundreds the built-in LLM settings into the shell surroundings. Within the very subsequent hands-on step, learners scan a malicious agent card with the LLM analyzer enabled.
supply ./lab-env.sh
a2a-scanner scan-card examples/malicious-agent-card.json --analyzers llm✅ Lab LLM settings loaded
Supplier: openai
Mannequin: gpt-4o
💡 Now you can run: a2a-scanner list-analyzers
Scanning agent card: Official GPT-4 Monetary Analyzer
Scan Outcomes for: Official GPT-4 Monetary Analyzer
Goal Kind: agent_card
Standing: accomplished
Analyzers: yara, heuristic, spec, endpoint, llm
Whole Findings: 8
description AGENT IMPERSONATION Agent falsely claims to be verified by OpenAI
description PROMPT INJECTION Agent description incorporates directions to disregard earlier directions
webhook_url SUSPICIOUS AGENT ENDPOINT Agent makes use of suspicious endpoints for information assortment 
That lab-env.sh step is the entire level: it preloads the managed lab LLM configuration into the terminal session, so the scanner can name the mannequin immediately with none handbook supplier setup. From the learner’s perspective, it feels virtually native, as a result of they supply one file and instantly begin utilizing LLM-backed evaluation from the command line.
How It Works
Why a proxy? The LLM Proxy abstracts a number of suppliers behind a single OpenAI-compatible endpoint. Learners write code towards one API—the proxy handles routing to Azure OpenAI or AWS Bedrock primarily based on the mannequin requested. This implies lab content material doesn’t break once we add suppliers or change backends.
Quota enforcement occurs on the proxy, not the supplier. Every request is validated towards the token’s remaining price range and request rely earlier than forwarding. When limits are hit, learners get a transparent error—not a shock invoice or silent failure.
Each request is tracked with consumer ID, lab ID, mannequin, and token utilization. This offers lab authors visibility into how learners work together with LLMs and helps us right-size quotas over time.
Fingers-On with AI Safety
The primary wave of labs on this infrastructure spans Cisco’s AI safety tooling:
- A2A Protocol Safety — built-in LLM settings are loaded throughout setup and used instantly within the first agent-card scanning workflow
- AI Protection — makes use of the identical managed LLM entry within the BarryBot utility workouts
- Ability Safety — makes use of the identical managed LLM entry within the first skill-scanning workflow
- MCP Safety — provides LLM-powered semantic evaluation to MCP server and gear scanning
- OpenClaw Safety (coming quickly) — validates the built-in lab LLM throughout setup and makes use of it within the first actual ZeroClaw smoke take a look at
These aren’t theoretical workouts. Learners are scanning reasonable malicious examples, testing reside safety workflows, and utilizing the identical Cisco AI safety tooling practitioners use within the discipline.
“We wished LLM entry to really feel like the remainder of Studying Labs: begin the lab, open the terminal, and the mannequin entry is already there. Learners get actual hands-on AI workflows with out chasing API keys, and we nonetheless preserve the controls we want round value, security, and abuse. I additionally preserve my very own working assortment of those labs at cs.co/aj.” — Barry Yuan
What’s Subsequent
We’re extending Studying Labs to help GPU-backed workloads utilizing NVIDIA time-slicing. This may let learners work hands-on with Cisco’s personal AI fashions—Basis-sec-8b for safety and the Deep Community Mannequin for networking—working regionally of their lab surroundings. For the technical particulars on how we’re constructing this, see our GPU infrastructure collection: Half 1 and Half 2.
Your suggestions shapes what we construct subsequent. Strive the labs and tell us what you’d prefer to see.