Most organizations get more value by focusing on current, low-cost, high-likelihood risks rather than rare or lab-only scenarios. Based on recent data points from the field, there are several areas to prioritize:
1. Exposed and over‑permissive AI infrastructure
- A 2026 scan found over 12,000 instances of a popular self‑hosted model server open to the internet with no authentication.
- Researchers also caught attackers running tools on exposed model servers, leaving the owners to pay the compute bill and risk data exposure.
What to do: Put authentication in front of every AI server, keep them off the public internet where possible, and patch them on your normal cycle.
2. Data leakage through AI tools
- Staff paste source code, contracts, and customer data into hosted AI services.
- Retrieval-augmented systems can surface documents that users were never meant to see if permissions aren’t enforced correctly.
What to do: Offer an approved, private way to use AI inside your environment (for example, with models running on VCF). Make sure retrieval respects existing document permissions.
3. AI hallucinations exploited by attackers
- In a study of 16 code models, researchers found 200,000+ invented package names.
- In 2024, a researcher uploaded an empty package under a model-invented name and it was downloaded 30,000+ times in three months.
What to do: Treat AI output as a starting point. Verify facts, URLs, and package names. Review AI-written code as you would code from a new teammate.
4. Misconfigured automation and “AI with too many keys”
- Agents with broad tool access can make rapid, large‑scale changes.
- Studies in 2026 showed intermediaries like routers between agents and models sometimes injecting malicious code or misusing credentials.
What to do: Separate duties, limit autonomy, and require human approval for irreversible actions. Test your kill switches so you can shut systems down without going through the AI itself.
5. Basic security hygiene—still critical with AI
More and more incidents involve AI speeding up attacks, not inventing new ones. In one case in 2026, hundreds of AI agents helped compromise at least 395 organizations in 48 countries, with 11 compromised in just 26 seconds.
What to do: Continue to:
- Patch internet-facing systems quickly.
- Segment networks, especially for IoT and smart devices.
- Use strong authentication and least privilege.
- Monitor for unusual traffic and large model/file transfers.
A practical rule of thumb: focus first on AI threats that have already happened, cost little for attackers to attempt, and target technology you actually run. In many cases, strengthening your existing security fundamentals will do more to protect your AI initiatives than chasing the latest movie-style scenario.