The safest way for a small team to use AI is to treat it as both a cybersecurity system and a data-processing service. Before anyone pastes business information into a chatbot or connects an AI agent to company systems, inventory the tools, classify the data, verify the provider’s terms, enforce ordinary account security, test AI-specific attacks, and define who can stop and recover the service. The checklist below follows guidance for small businesses from the Australian Cyber Security Centre (ACSC), the U.S. Federal Trade Commission (FTC), NIST, OWASP and CISA.
What an AI security checklist must protect
AI risk is not limited to whether a model gives a wrong answer. A tool can expose confidential inputs, retain prompts, reveal information through an output, or use a connected account to change records and send messages. The consequences depend on the data, integrations and authority you grant it.
As an Amazon Associate I earn from qualifying purchases.
Use two questions for every proposed use: What information can this system receive or retrieve? and What can it do after it receives it? A text-only assistant that drafts an internal outline has a different risk profile from an agent that can read a customer database, create invoices or send email. The controls should become stricter as sensitivity, connectivity and autonomy increase.
The ACSC’s 2025 guide for small business specifically recommends checking what a service collects, where it stores information, who owns it, whether it may use data to train models, how outputs are checked, and how incidents are handled. Those questions form the data and vendor parts of this checklist.
#1 Best Overall
- 【Flexible Port Configuration】1 Gigabit SFP WAN Port + 1 Gigabit WAN Port + 2 Gigabit WAN/LAN Ports plus1 Gigabit LAN Port. Up to four WAN ports optimize bandwidth usage through one device.
- 【Increased Network Capacity】Maximum number of associated client devices – 150,000. Maximum number of clients – Up to 700.
- 【Integrated into Omada SDN】Omada’s Software Defined Networking (SDN) platform integrates network devices including gateways, access points & switches with multiple control options offered – Omada Hardware controller, Omada Software Controller or Omada cloud-based controller(Contact TP-Link for Cloud-Based Controller Plan Details). Standalone mode also applies.
- 【Cloud Access】Remote Cloud access and Omada app brings centralized cloud management of the whole network from different sites—all controlled from a single interface anywhere, anytime.
- 【SDN Compatibility】For SDN usage, make sure your devices/controllers are either equipped with or can be upgraded to SDN version. SDN controllers work only with SDN Gateways, Access Points & Switches. Non-SDN controllers work only with non-SDN APs. For devices that are compatible with SDN firmware, please visit TP-Link website.
1. Make an inventory before approving another AI tool
You cannot secure tools that nobody knows are in use. Create a living inventory, even if it starts as a spreadsheet. Record each application, model or API, the people and service accounts using it, its business purpose, and every integration.
- Application and model: name, edition, workspace or API account, and the date it was approved.
- Users: employees, contractors, shared accounts and automated service accounts.
- Inputs: prompts, uploaded files, email, tickets, source code, customer records or sensor data.
- Connected systems: storage, identity provider, CRM, accounting, code repositories, messaging and web tools.
- Outputs and actions: draft text, decisions, database writes, payments, file changes or outbound communications.
- Data location and retention: what the provider states about storage regions, deletion and backups.
- Owner: one person responsible for access reviews, supplier changes and incident contacts.
Include personal subscriptions and browser extensions. A free trial used with a company account can create an unmanaged copy of business data. Ask staff to report new tools rather than trying to block every experiment; visibility is more useful than a policy nobody follows.
2. Classify information before anyone shares it
Write a plain-language rule for what may be entered, what requires approval and what is prohibited. The provider’s marketing name is not evidence that a plan is suitable for confidential data. Confirm collection, storage, ownership, deletion and model-training terms for the exact edition and region your team will use.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minute| Information type | Default handling | Checks or safeguards |
|---|---|---|
| Public material | Permitted for approved tools | Still remove hidden metadata and verify the output before publication. |
| Internal, low-sensitivity work | Permitted only in an inventoried service | Use named accounts, MFA and the provider’s stated retention and deletion controls. |
| Customer or employee personal information | Approval required; often keep out of general-purpose tools | Confirm lawful handling, access limits, storage location, retention and whether training use is disabled or contractually excluded. |
| Credentials, API keys and security secrets | Never paste into a prompt or upload | Store in a secrets manager and rotate immediately if exposed. |
| Financial records, unreleased plans or regulated data | Use only a specifically approved workflow | Document the business need, contractual protections, encryption, deletion and human review. |
Ownership matters as much as confidentiality. Ask whether your organization retains rights to inputs and outputs, whether the provider can use them to improve models, which subprocessors receive them, and how a deletion request works. Keep a copy of the relevant terms and privacy documentation because services change. The ACSC checklist is a useful set of questions for this review: Artificial intelligence for small business: Managing cyber security risks.
A practical “safe to share” decision
- Remove secrets, direct identifiers and unnecessary fields.
- Check the inventory for the approved service and its current owner.
- Read the service’s collection, retention, deletion and training-use statements.
- Obtain the designated approver’s decision for personal, financial, regulated or strategically sensitive information.
- Record what was shared and for which business purpose when the data is sensitive.
3. Review the provider and its service terms
A security logo or a promise of “enterprise AI” is not a complete assessment. Review the contract, privacy notice, security documentation and support process for the actual product and plan.
| Review area | Questions for a small team |
|---|---|
| Data handling | What is collected from prompts, files, telemetry and logs? Where is it stored? How long is it retained? Can the provider or a subprocessor use it for model training? |
| Ownership and deletion | Who owns inputs and outputs? Can data be exported and deleted, including backups? What happens when the account is closed? |
| Security evidence | What independent assurance, penetration-test summary, vulnerability process or encryption details are available? |
| Dependencies | Which cloud hosts, model suppliers, plug-ins and connectors can access the data, and how are they changed or notified? |
| Access and administration | Does the service support MFA, roles, single sign-on, audit logs, session control and rapid account revocation? |
| Response and availability | How will the provider notify you of a breach or material change? What support channel handles an urgent compromise, and how can work continue during an outage? |
Revisit the review when the provider changes the model, plan, retention policy, connector or geographic processing location, and whenever your team expands the use case. Keep a short written decision explaining why the service is acceptable and what restrictions apply.
4. Apply ordinary account and device security
AI accounts are still accounts. The FTC’s Cybersecurity for Small Business guidance recommends baseline measures that protect AI workspaces and the systems connected to them: multifactor authentication (MFA), restricted access, encryption, backups, monitoring and incident planning. Its page describes the latest CSF 2.0 as “free, voluntary, and flexible.”
Quick wins for a faster PC:
Clear out junk files and repair common Windows errorsFree Scan →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →Repair Windows errors before they cause bigger problemsFix Now →- Require MFA. Enable it for every AI, email, identity and administrator account. A FIDO2 security key can be one implementation option when it is compatible with your identity provider; document spare-key storage and account recovery before enforcing it.
- Use individual accounts and least privilege. Remove shared logins, assign roles, and give a user or integration only the data and actions needed for the task.
- Change defaults and secure recovery. Replace default passwords, protect recovery email and phone numbers, and review active sessions and API tokens.
- Update endpoints and integrations. Keep operating systems, browsers, extensions, connectors and security software current. Remove unused plug-ins and revoke stale tokens.
- Encrypt and back up important data. Protect data in transit and at rest where the service supports it, and maintain tested backups that an AI account cannot delete.
- Monitor access. Review sign-in, token, prompt, file and tool-call logs where available. Alert on unfamiliar locations, mass downloads, privilege changes and unusual automated activity.
Apply these controls to the identity that owns an integration, not only to the employee who configured it. A compromised service account can bypass a well-trained user if it has broad, long-lived permissions.
Rank #2
- Built on a purposed-built secure processor, this compact network firewall delivers the highest level of security performance and energy efficiency in its class – 2.25 Gbps IPS throughput | 1.1 Gbps threat protection | 1.3 Gbps SSL Inspection throughput.
- User-friendly management console gives you centralized visibility and simplifies policy enforcement across your network. Its zero-touch deployment helps you optimize your onboarding experience.
- Compact and fanless design equipped with 5 GE RJ45 ports (1 WAN port and 4 internal ports).
- Fortinet is the most deployed and trusted firewall from businesses worldwide with 99.98% security effectiveness, surpassing competition. Fortinet is the only vendor recognized as a firewall leader 13 consecutive years by Gartner.
5. Train staff and check every consequential output
Training should explain the data-classification rule, approved tools, prohibited inputs, account-reporting route and examples of prompt injection. The ACSC guide names Cyber Wardens Level Two – Safe AI for Small Business as an example of staff education; use an equivalent course or an internal briefing suited to your jurisdiction and work.
Make reporting easy: staff should know whom to contact if they paste a secret, receive a suspicious instruction, see an unexpected file change or discover a new AI service. Treat a report as a security signal, not a disciplinary event, so problems surface quickly.
Require a human review before AI-generated content is sent externally or used for a material decision. Check factual claims against primary records, calculations, citations, names, dates, permissions and confidential details. For customer, employment, health, safety, legal, financial or access decisions, define an accountable human decision-maker and retain the evidence used to approve the result. Fluency is not proof of accuracy, and a model can confidently follow a malicious instruction embedded in a document.
Do these 3 things before closing this tab:
1Repair Windows errors before they cause bigger problems2Fix the driver behind crashes, sound loss and screen glitches3Clear out junk files and repair common Windows errors6. Test AI-specific attack paths
Ordinary vulnerability scanning does not reveal every AI failure. Use the categories in the OWASP AI Security Verification Standard to structure a review: input validation, configuration and lifecycle, identity and access, supply chain, output controls, memory and vector databases, agents, orchestration, adversarial robustness, monitoring and logging.
Prompt injection and untrusted content
Test whether instructions hidden in a web page, document, email or retrieved record can override the system’s intended rules. NIST’s Generative AI Profile (NIST AI 600-1) explicitly recommends red-teaming for prompt injection, adversarial prompts, data poisoning, membership inference, model extraction and related attacks. Separate untrusted content from system instructions, restrict what retrieved text can trigger, and require confirmation before a tool call with side effects.
Disclosure and output abuse
Ask the system for secrets, other users’ records and hidden instructions using realistic test accounts. Check whether outputs include personal data, credentials, source documents or unsafe code. Add output filters and approval gates where exposure or harm would be consequential, while recognizing that filters are not a substitute for access control.
Integration and supply-chain failures
Review connector scopes, plug-in provenance, dependency updates, webhook authentication and the data returned by retrieval systems. Test what happens when a connected service is unavailable, returns hostile content or changes its schema. Keep high-risk connectors disabled until an owner has approved them.
Free tools Windows power users keep installed
One-click scans. No signup required.
Measure and repeat
Record the test case, expected behavior, observed result, severity, owner and fix date. Retest after a model, prompt, connector, permission, data source or workflow changes. NIST’s recommendation is to assess resilience, not to assume that one successful test proves safety.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.7. Monitor use and reassess as the system changes
Choose a small set of signals your team can actually review: new users and tokens, unusual prompt volume, bulk exports, failed authorization, unexpected tool calls, sensitive-data matches, model or connector changes, and overrides of human approval. Set a review schedule—monthly for active integrations is a reasonable starting point—and perform an additional review after a significant change or incident.
Keep logs long enough to investigate an event, subject to your privacy and retention obligations. Limit who can read them because prompts and tool arguments may contain sensitive information. Document a clear owner for each alert; an unassigned dashboard is not monitoring.
8. Prepare an incident plan a small team can execute
Write a short playbook before a compromise. It should identify the person who can disable the AI account or connector, the person who contacts the provider, the person who preserves evidence, and the person who coordinates customer, employee, regulator or law-enforcement notifications when required by location, sector or contract.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →- Contain: revoke tokens, disable connectors, suspend affected accounts and isolate connected devices without destroying evidence.
- Assess: determine what prompts, files, records or actions were exposed or changed, using provider and identity logs.
- Recover: rotate secrets, restore trusted data from backups, rebuild permissions and re-enable only the minimum required functions.
- Communicate: use the provider’s urgent channel and follow applicable contractual and legal notification duties.
- Learn: document the root cause, update the inventory and data rules, and retest the corrected workflow.
Run a tabletop exercise with the actual people and contact details. Include the case where the AI service is unavailable, because business continuity may require a manual process rather than a technical fix.
9. Add tighter guardrails to AI agents
An agent combines a model with tools, memory and authority. CISA and partner agencies’ May 1, 2026 announcement on careful adoption of agentic AI services recommends limiting autonomy and access, using layered defenses, strong identity, oversight, threat modeling, monitoring and regular assessments: CISA announcement.
- Give the agent a separate identity and the narrowest possible role; do not reuse a human administrator account.
- Allow only the specific tools, records, fields, destinations and time window needed for the task.
- Make read-only access the default. Require explicit human approval for payments, deletion, permission changes, legal or employment decisions, and external messages.
- Use rate limits, transaction limits, destination allowlists and reversible actions where possible.
- Log every tool call, argument, result, approval and refusal, and alert on unusual sequences.
- Assume retrieved documents and tool responses are untrusted input; prevent them from rewriting the agent’s governing instructions.
- Provide a tested kill switch and a manual fallback for essential work.
Before production, threat-model the agent’s data paths and run abuse tests with a nonproduction account. Recheck permissions whenever a new tool, memory store, workflow or model is added.
How to prioritize when nobody owns security full time
Start with the controls that reduce the largest exposure quickly: inventory every tool, prohibit secrets and unapproved personal data in prompts, turn on MFA, remove shared and excessive access, and name an incident owner. Next, review the terms for the few services handling sensitive information and add human approval to consequential actions. Then schedule AI red-team tests, logging reviews and recovery exercises. A short, maintained process is safer than a comprehensive policy that no one can operate.
Use the checklist as a decision record: each service should have an owner, approved data classes, allowed integrations, required human checks, monitoring responsibilities and a documented response path. If you cannot explain those six items, do not connect the service to sensitive data or high-impact actions yet.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
