AI Is Coming: Meet the Startups Building Cyber Defenses for the Age of AI

1. Visibility & Governance

The first step in securing AI is gaining visibility into how employees interact with AI tools: identifying which models they use, what data they input, and what they receive in return. Once CISOs have a clear view of AI usage within the organization, they can implement control measures (also known as governance) to manage AI access and usage. Governance applies to both third-party AI tools (like ChatGPT) and internally developed AI tools (such as an in-house chatbot).

Internal AI Models

For internally built AI, Role-Based Access Control (RBAC) is essential to restrict access to sensitive data. This is a security method that gives access to systems, applications, and data based on a user’s role, ensuring resource access is only given to those who need it for their job functions.

For example, if a company builds an AI-powered chatbot trained on internal Slack conversations, it’s crucial to ensure the security of the data layer that contains the private Slack conversations. This involves attaching metadata to the chat data so that only authorized users can access specific information – you don’t want just anyone viewing private Slack conversations.

A key innovation in this space is policy-as-code, which allows teams to define security policies using code rather than relying on manual processes (like writing directly to configuration files). Policy-as-code frameworks can include rules, conditions, and automated protections — such as blocking unauthorized access when breaches occur. CISOs can use programming languages like Python or Rego to define policies, making security faster and less error-prone.

Backward compatibility is also essential. New AI security tools must seamlessly integrate with existing enterprise system like Salesforce to prevent accidental data leaks and maintain consistent access control.

External AI Services

For third-party tools like ChatGPT, similar policy enforcement principles can apply. CISOs can set controls specifying which tools employees are allowed to use, define the types of data that can be input to the tools, and establish safeguards to prevent data leaks, often using policy-as-code or comparable frameworks.

While most AI security tools today focus on human interactions with AI, the scope is quickly evolving. APIs and AI agents are starting to access AI models too, so very soon we expect AI security solutions to secure these non-human AI interactions as well.

2. Pre-Deployment Testing

The next layer of AI security focuses on selecting the right AI model and preparing it for deployment within your organization. As of 2025, the landscape of AI models is rapidly expanding, with platforms like Hugging Face offering 1.5M models and 40,000 new models added every week. However, choosing the right model isn’t just about performance; it’s also about evaluating factors like bias and security.

Most AI models are trained on datasets sourced from the internet, making them susceptible to biases and misinformation. These biases can result in reputational damage and even legal repercussions. A prominent example occurred in 2024 when Google’s Gemini model sparked controversy by generating racially diverse images, even when users explicitly requested images of a specific race.

To mitigate such risks, many enterprises are investing in AI auditing and compliance. Similar to how companies use tools like Drata and Vanta to achieve SOC 2 certification for security and privacy, organizations are increasingly adopting AI model certification solutions to ensure their AI models are free from bias and toxicity. Additionally, businesses are partnering with vendors that assess AI models for security risks and ethical concerns, providing documentation for regulators and customers.

Another emerging trend is firms using AI to detect high-risk employee behaviors that could compromise security. For example, Dune Security’s platform identifies risky actions, such as clicking on phishing links or visiting unsecured websites, and flags them as potential threats.

Continuous Vulnerability Scanning

Continuous vulnerability scanning is gaining momentum as a component of AI security. Unlike traditional software, AI models cannot be effectively scanned using existing Endpoint Detection and Response (EDR) tools, which often fail to detect AI-specific threats such as payload attacks, prompt engineering, and exposure of personally identifiable information (PII).

For example, the AI model DeepSeek has demonstrated several vulnerabilities identified by AI security startups like Holistic AI and Hidden Layer. Certain prompts can inadvertently cause DeepSeek to reveal its training data, while structured prompts can trap the model in an endless loop — a phenomenon known as Adversarial Chain of Thought. A red-teaming report by Enkrypt AI revealed that DeepSeek was 11 times more likely to generate harmful content, 4 times more toxic, and 4.5 times more likely to produce insecure code compared to leading models like GPT-4 and Claude-3. Consequently, new tools are emerging that provide continuous scanning of AI models, identifying vulnerabilities and highlighting areas that require remediation.

Currently, most solutions focus solely on protecting and identifying vulnerabilities within the AI model itself. However, the next wave of innovation is expected to go beyond model protection, aiming to secure the entire AI ecosystem. This would include the application, its data flows, containerized services, vector databases, and other integrated infrastructure components.

3. Post Deployment Defense

So, you’ve chosen your model, scanned for vulnerabilities, tested its performance, and are ready to deploy it to real users. But you’re not done yet. Once an AI system goes live, it becomes a prime target for attackers seeking to exploit its weaknesses in various ways.

Let’s break down a key list of attack vectors, though this list is not exhaustive.

Many solutions are emerging to defend against these attack vectors. One promising approach is the AI firewall, which acts like a security guard for your AI models.

AI Firewall

Just as Web Application Firewalls (WAFs) monitor and protect web applications from harmful traffic, an AI firewall monitors and filters the data flowing in and out of an AI model to detect suspicious activity or malicious inputs that could compromise the model. One example is Robust Intelligence’s AI Firewall, which validates model inputs and outputs in real time, protecting against various threats, including prompt injection, PII extraction, and hallucination. Robust Intelligence’s AI firewall is integrated with MongoDB Atlas Vector Search, allowing MongoDB’s 40,000+ customers to use any commercial or open-source AI solutions safely.

Robust Intelligence’s AI security solutions have gained significant traction, with its products being used by JPMorgan Chase, Expedia, Deloitte, Cisco, and even the U.S. Department of Defense. The adoption highlights the growing need for reliable AI security solutions today, and we expect the demand to grow.