YOUR AI PLATFORM NEEDS MORE THAN INTELLIGENCE. IT NEEDS PROTECTION.

Our team will help identify:

• Who owns the main account
• Who requires access
• What each person needs to do
• Which accounts have administrator authority
• Whether former employees or contractors still have access
• When temporary access should expire

Every authorized person should have an individual account. Shared administrator accounts should not be used.

A password should not be the only protection for an important business account. During onboarding, the team should:

• Require multifactor authentication for administrators.
• Help the client activate MFA on supported platforms.
• Prefer passkeys, security keys, or authenticator applications.
• Securely store emergency recovery codes.
• Explain that verification codes must never be shared.
• Confirm that recovery email addresses and phone numbers are current.

CISA recommends MFA because it requires an additional method of verifying a person’s identity. Stronger, phishing resistant options should be used when available.

Before granting access to a person, contractor, developer, or outside company, confirm:

• Full name
• Company or organization
• Business role
• Reason access is needed
• Person who approved the request
• Systems they are allowed to use
• Access expiration date

Sensitive access requests should be confirmed through a second trusted communication method. Do not rely only on an unexpected email or text message.

Not everyone needs full control of the platform. Create separate roles such as:

• Owner
• Administrator
• Developer
• Sales representative
• Customer support
• Content editor
• Read only user

Each person should receive only the access needed to perform their assigned responsibilities. Administrator access must be limited.

Before connecting any AI tool, plugin, application, or automation, document:

• Official product and company name
• Verified developer or publisher
• Information the tool can access
• Actions the tool can perform
• Permissions being requested
• Information storage practices
• Whether customer information may be used for model training
• Method for removing the connection
• Person responsible for managing the tool

Unknown or unverified tools must not be connected. OWASP identifies prompt injection, sensitive information disclosure, supply chain vulnerabilities, and excessive AI authority as significant risks for AI applications.

The tech team must:

• Use an approved password manager.
• Store API keys in an encrypted secrets manager.
• Never place passwords or API keys in website code.
• Never send credentials through unprotected email or text messages.
• Use different credentials for testing and live systems.
• Replace credentials when team members leave.
• Deactivate old or unused keys.
• Limit every key to the permissions it needs.

Sensitive information should not be entered into an AI system unless the platform is approved to receive it and the business has a valid reason to use it. Protected information may include:

• Social Security numbers
• Medical information
• Financial records
• Payment information
• Passwords and verification codes
• Driver’s license information
• Confidential agreements
• Employee records
• Private customer communications

The platform should collect only the information needed to serve the customer.

AI should not be given unlimited authority. Require human approval before AI can:

• Send payments or refunds
• Delete customer information
• Change user permissions
• Sign or accept agreements
• Publish sensitive content
• Transfer confidential files
• Send legally significant communications
• Make important employment or customer decisions

OWASP warns that excessive AI authority can cause harmful actions when a system receives too much functionality, permission, or independence.

Track important events, including:

• Successful and failed sign ins
• New user accounts
• Permission changes
• Administrator activity
• Connected and disconnected tools
• Data downloads and exports
• API key activity
• Automation changes
• Unusual AI usage
• Unexpected platform costs

Create alerts for repeated failed sign ins, new administrators, large data exports, suspicious locations, and unexpected permission changes.

Security responsibilities continue after launch. Recommended schedule:

• Daily: Review critical security alerts.
• Weekly: Review unusual activity and failed sign ins.
• Monthly: Review users, tools, permissions, and software updates.
• Quarterly: Remove unused accounts, test backups, and complete a security review.
• Immediately: Remove access when a team member, contractor, or vendor leaves.

NIST recommends managing AI risks across the complete AI system life cycle through governance, risk identification, measurement, and ongoing management.