What Is Kinovi AI and How Does It Work?

A startup checklist for AI tool security: risks, vendor questions, contracts and light governance.

Most startups adopt AI tools before they have any process for checking them. A founder pastes customer data into a free chatbot, an engineer connects a coding assistant to the repo, and sensitive information starts flowing to systems nobody has reviewed.
Standard SaaS vendor checklists miss the risks that matter most with AI: whether your inputs train someone else's model, where prompts actually go, and which third parties sit behind the interface. This guide gives you a practical framework for evaluating AI tool security without a dedicated security team: the core risks, what to check, what to ask vendors, and a light governance routine that fits an early-stage company.
To evaluate an AI tool's security, focus on four areas: data handling, model training policies, vendor compliance and how the tool connects to your systems. A standard SaaS review, such as checking for a SOC 2 badge and skimming the privacy policy, misses risks that are specific to AI.
AI tools raise questions traditional software doesn't. Where do prompts and file uploads go? Does the vendor use your inputs to train its models? Which model providers sit behind the interface?
A familiar checklist works for a file-sharing app. An AI tool that reads customer contracts, strategy documents or source code needs a different lens. The rest of this guide provides it.
Each risk below becomes a check during vendor review. Where it fits, we note the matching entry in the OWASP Top 10 for LLM Applications 2025, the most widely used list of security risks for applications built on large language models.
Text, files and images you send to an AI tool may be stored, logged or visible to the vendor's staff. Any prompt with customer information, financial data, contracts or strategy becomes a possible exposure point. A chat window feels temporary, but often it isn't: retention periods vary widely, and some vendors keep inputs for weeks or longer unless you change a setting. OWASP lists this as LLM02: Sensitive Information Disclosure.
Some AI tools use prompts, documents or outputs to train or improve their models. If they do, your startup's information could shape future outputs. Look for a clear statement in the terms or Data Processing Agreement that your data is excluded from training, or a setting that turns training off.
Consumer plans often leave this choice to you. Anthropic, for example, lets Claude users choose whether their chats are used to improve its models and change that choice at any time in Privacy Settings. Business plans tend to start from the safer default: ChatGPT's Business plan does not train on business data by default, according to AITL's pricing check.
Prompt injection happens when hidden or malicious instructions inside an input change how the model behaves. It matters most when you use AI in customer-facing workflows, support systems or automations that read outside content, such as emails, documents or web pages. It is first on the OWASP list as LLM01: Prompt Injection.
AI agents that connect to Slack, email, CRMs, calendars or code repositories may ask for broad permissions. If those permissions are too wide, the tool can read or change more data than you intended. Review exactly what a tool requests before you connect it; read-only access is usually safer than admin rights. OWASP calls this LLM06: Excessive Agency.
The AI tool you buy may pass your data to a model provider such as OpenAI, Anthropic or another hosted service. You need to know who else processes your data, under what terms, and whether they appear on a published sub-processor list. A product can have a clean privacy policy while the model behind it follows different retention or training rules. OWASP covers this under LLM03: Supply Chain.
Startups often start on free or consumer plans and later wonder when to upgrade. The difference matters because consumer plans usually lack the controls that protect business data.
| Factor | Consumer plans | Business and enterprise plans |
|---|---|---|
| Data retention | Set by the vendor, often long or unclear | Configurable, sometimes with zero-retention options |
| Training on your data | Often on by default, with a setting to turn it off on some tools | Usually off by default or excluded by contract |
| Admin controls | None | Roles, permissions and audit logs |
| Data Processing Agreement | Rare | Standard |
| Sign-in and identity | Email login | SAML SSO, SCIM and enforced MFA |
ChatGPT shows the difference in practice. Its Business plan costs $25 per user per month, or $20 billed yearly, for teams of 2 or more, and adds admin controls, SAML SSO and MFA, with no training on business data by default.
Free plans can be fine for low-risk brainstorming and internal drafts. For customer data, financial information, legal material or proprietary code, use a plan with explicit controls.
Use these criteria for every AI tool you consider.
Send these before you sign up or buy:
The questionnaire is a start. Many problems only show up in the contract.
A Data Processing Agreement (DPA) explains how the vendor processes data on your behalf. Check that the DPA or a linked policy says clearly whether your data is excluded from model training.
The sub-processor list shows who else touches your data. Check that the vendor will notify you before adding a new sub-processor, rather than quietly updating a web page.
Check how quickly the vendor commits to telling you about a security incident. Under Article 33 of the GDPR, a company that controls personal data must report a breach to the supervisory authority within 72 hours where feasible, and a vendor processing data for it must notify it "without undue delay". Because your own 72-hour clock depends on the vendor, look for a specific notification window in the contract.
Shadow AI means AI tools people use without approval. It happens even in small teams. Practical ways to find it:
The goal isn't to punish anyone. It is to know which tools are already in use so you can review them.
The U.S. National Institute of Standards and Technology released the AI Risk Management Framework on January 26, 2023. It is voluntary and organizes AI risk work into four functions: Govern, Map, Measure and Manage. It is useful for structuring your risk documentation, especially when investors or enterprise customers ask how you manage AI risk.
The OWASP list names the most common security risks in applications built on large language models, including Prompt Injection (LLM01), Sensitive Information Disclosure (LLM02), Supply Chain (LLM03), Improper Output Handling (LLM05) and Excessive Agency (LLM06). Use it when your team builds on AI APIs or puts AI into customer-facing products.
AI tool review isn't a one-time task. A minimal routine keeps oversight in place without heavy process.
Most startups don't have security staff, and this framework is built for that. Use the checklist and vendor questions above as your template. Bring in a part-time security advisor for one-off help on high-risk tools. Review tools that touch customer data or source code first, and lower-risk productivity tools later.
A curated directory can also cut the first round of research. At AITL, we read each tool's official website, pricing, documentation and terms, and try it on a real task before listing it, as described on our How We Curate page. That doesn't replace your own security review, but it gives you a shortlist of tools with public pricing and documentation to start from. You can browse the full directory or put two or three side by side on the Compare page.
The main risks are data leakage through prompts and uploads, training on your inputs, prompt injection, over-permissioned agents and integrations, and data passing to third-party model providers. The OWASP Top 10 for LLM Applications 2025 lists prompt injection first.
Ask for the vendor's SOC 2 Type II report or a similar attestation, read the Data Processing Agreement, confirm how training on your data is handled, check the sub-processor list, and test the tool with non-sensitive data before you use real business information.
Not fully. A SOC 2 report covers general security controls, but not AI-specific issues such as prompt injection or training on customer data. Ask AI-specific questions on top of the SOC 2 report.
Free and consumer plans usually lack admin controls, Data Processing Agreements and business-grade data terms, and some use your chats for training unless you turn it off. They can be fine for low-risk internal work. For customer data, financial information or proprietary code, use a business plan with explicit controls.
It is a voluntary framework from the U.S. National Institute of Standards and Technology, released on January 26, 2023. It organizes AI risk management into four functions: Govern, Map, Measure and Manage.
Founder and editor of AITL. Researches, tests and curates the AI tools listed in the directory.
Every week we combine 2–4 tools from AITL into a step-by-step workflow you can try in minutes.
Plus the newest tools we listed.