AI agents are everywhere. Tools like OpenClaw (which has racked up over 192,000 GitHub stars) promise to transform how businesses operate - autonomous AI that can work across Slack, Teams, WhatsApp, and Discord, with access to nearly 4,000 pre-built skills.

But here’s what most vendors won’t tell you upfront: deploying AI agent platforms in a business environment is one of the riskiest IT decisions you can make right now.

The OpenClaw Reality Check

OpenClaw is open-source, powerful, and popular. It’s also a security nightmare if you don’t know what you’re doing.

Recent security audits revealed that 36.82% of skills available on ClawHub (OpenClaw’s skill marketplace) contain security flaws. We’re not talking about minor bugs - these are vulnerabilities that could expose your business data, give attackers access to your systems, or leak sensitive customer information.

Worse still, researchers traced 341 confirmed malicious skills back to a single coordinated campaign. Someone’s actively building and distributing compromised AI agent skills, and they’re doing it at scale.

Add to that: over 30,000 OpenClaw instances are currently exposed to the internet because the default configuration makes them publicly accessible. Most businesses don’t even realise their AI agents are broadcasting to the world.

Why SMBs Are the Primary Target

Small and medium businesses are particularly vulnerable here because you’re often first to adopt new technology but last to get enterprise-grade security guidance.

AI agent platforms represent a new attack vector that most SMBs aren’t prepared for. Your traditional security controls - firewalls, antivirus, email filtering - weren’t designed to monitor autonomous AI making decisions and taking actions on behalf of your business.

Think about what an AI agent can do:

• Access your email and messaging platforms
• Read and write files in cloud storage
• Make API calls to third-party services
• Process customer data
• Execute code based on external inputs

Now imagine that AI agent running a skill with a backdoor installed by an attacker. That’s not a theoretical risk. It’s happening right now.

The Security Audit Checklist

Before you deploy any AI agent platform - OpenClaw or otherwise - here’s what you need to verify:

1. Skill Provenance and Vetting

What to check:

• Where do the skills come from? Who built them?
• Is there a formal security review process before skills are published?
• Can you audit the source code of every skill you’re deploying?
• How quickly are vulnerabilities patched when they’re discovered?

Red flags:

• Skills with obfuscated code or binary-only distributions
• No clear maintainer or organisation behind the skill
• Requests for excessive permissions that don’t match the skill’s stated function
• Poor or missing documentation

2. Network Exposure and Access Controls

What to check:

• Is the AI agent platform exposed to the internet by default?
• Can you restrict access to specific IP ranges or VPN-only?
• Does it support proper authentication (not just API keys)?
• Can you implement role-based access control?

Red flags:

• Default configurations that allow public access
• Hard-coded credentials or API keys in configuration files
• No support for multi-factor authentication
• Limited ability to segregate different teams or projects

3. Data Handling and Privacy

What to check:

• Where is data processed and stored?
• Does the platform send data to external services without disclosure?
• Can you control which skills have access to sensitive data?
• Is there audit logging of all data access?

Red flags:

• Unclear data residency (especially important for Australian businesses under Privacy Act obligations)
• Skills that phone home to unknown servers
• No data classification or access control mechanisms
• Missing or inadequate logging

4. Supply Chain Security

What to check:

• How do skill updates get distributed?
• Can malicious actors inject code into the update process?
• Is there integrity verification (digital signatures, checksums)?
• Can you freeze skill versions and control updates?

Red flags:

• Automatic updates with no review or approval process
• No verification of update authenticity
• Dependencies on unmaintained or abandoned packages
• Unclear trust chain from skill author to your deployment

5. Monitoring and Incident Response

What to check:

• Can you see what actions the AI agents are taking in real-time?
• Are there alerts for suspicious behaviour?
• Can you quickly disable a compromised skill?
• Is there an audit trail for forensics?

Red flags:

• Limited or no logging of AI agent actions
• No way to roll back or quarantine problematic skills
• Inadequate monitoring of API calls and data access
• No integration with your existing security information and event management (SIEM) tools

Practical Deployment Advice

If you’re serious about deploying AI agents securely, here’s what actually works:

Start with a private, isolated environment. Don’t connect AI agents directly to production systems until you’ve tested them thoroughly. Use separate accounts, dummy data, and sandboxed environments.

Vet every skill manually. Yes, it’s tedious. But spending an hour reviewing a skill’s code and permissions is better than spending weeks recovering from a breach. If you can’t read code, hire someone who can - even for a few hours.

Implement defence in depth. Don’t rely solely on the AI platform’s security. Add network segmentation, application-level firewalls, and monitoring tools. Use the principle of least privilege - give AI agents the minimum access they need, nothing more.

Choose Australian-hosted infrastructure where possible. This matters for privacy compliance, but it also reduces latency and gives you clearer legal recourse if something goes wrong. If you’re using AI consultants Sydney or similar services, verify their hosting arrangements.

Consider managed services. For many SMBs, a managed OpenClaw deployment with security hardening and pre-audited skills is more cost-effective than building internal expertise. You’re paying for someone else to do the security reviews, monitoring, and incident response.

What the ACSC and OWASP Say

The Australian Cyber Security Centre hasn’t yet issued specific guidance on AI agent platforms (they’re working on it), but their existing AI security guidance makes several points relevant here:

• Treat AI systems as high-risk attack surfaces
• Implement robust input validation (AI agents can be manipulated through prompt injection)
• Maintain human oversight of critical actions
• Ensure transparency and auditability

The OWASP Top 10 for LLM Applications (which applies to AI agents) highlights supply chain vulnerabilities as a critical risk. Their recommendation: “Just because a model or plugin is popular doesn’t mean it’s secure.”

The Bottom Line

AI agents like OpenClaw offer genuine business value. They’re not snake oil. But they’re also not ready for careless deployment.

If you’re a small or medium Australian business looking at AI agent platforms, don’t rush. The technology will still be there in six months, but your business might not if you deploy insecurely today.

Do your security audit first. Ask hard questions. Get professional advice if you need it. And remember: the goal isn’t to avoid AI agents entirely - it’s to deploy them in a way that doesn’t expose your business to preventable risks.

Because the attackers are already building compromised skills and waiting for businesses to install them. Don’t make it easy.