5.2. AI Standarts

Responsible AI and Compliance with AI Standards

1. Our Approach to AI Governance

TozheBot is designed with the principles of responsible artificial intelligence, information security, personal data protection, and AI risk management in mind.

Our platform enables users to build and operate AI agents, connect tools and knowledge bases, and integrate external services. Because agents may process user data, access corporate information, and perform actions through connected tools, AI security must be addressed not only at the language model level but across the entire platform.

We follow a layered approach that includes:

• Access control across users, projects, and agents.
• Permission management for agent tools and knowledge bases.
• Protection against potentially malicious instructions and attempts to manipulate AI models.
• Cryptographic protection for configured data fields.
• Security event monitoring.
• Auditing of supported personal data operations.
• Ongoing development of risk assessment and continuous improvement processes.

These mechanisms provide a technical foundation for responsible AI management. However, technical controls are only one part of a broader governance system. Comprehensive alignment with applicable standards also requires documented procedures, clearly assigned responsibilities, risk assessments, and evidence that controls operate effectively.

2. ISO/IEC 42001: Artificial Intelligence Management System

ISO/IEC 42001:2023 is an international standard that specifies requirements for establishing, implementing, maintaining, and continually improving an Artificial Intelligence Management System (AIMS).

The standard addresses AI governance at the organizational level, from setting objectives and assessing risks to controlling operations and improving processes. It applies to organizations that develop, provide, or use AI systems.

Official standard: https://www.iso.org/standard/42001.html

TozheBot is developing its architecture and security practices with several areas relevant to ISO/IEC 42001 in mind.

AI risk management

AI agents may encounter prompt injection, attempts to expose confidential information, incorrect outputs, and unwanted actions performed through connected tools.

TozheBot includes technical mechanisms designed to help identify and mitigate certain classes of these risks:

• The OGL-Mini AI security layer analyzes messages for indicators of potentially unsafe instructions and certain types of personal data.
• Access controls restrict agents' access to tools and knowledge bases.
• Security events provide a basis for recording and investigating suspicious activity.
• Personal data auditing helps track selected operations involving personal information.

To support a more mature ISO/IEC 42001-aligned management system, these controls should be complemented by systematic risk assessments, documented decisions, assigned risk owners, and periodic reviews of control effectiveness.

Accountability and oversight

AI security should not depend exclusively on the behavior of a language model. TozheBot therefore provides platform-level controls, including project isolation and permissions governing the resources available to agents.

An organizational AI management system should additionally define:

• Who is responsible for assessing and accepting AI-related risks.
• Who approves the deployment of new features and integrations.
• Who investigates and manages incidents.
• How changes and exceptions to security policies are documented.
• How the effectiveness of safeguards is reviewed.

Monitoring and continual improvement

Security events and audit logs can help identify recurring issues, investigate incidents, and improve protective measures.

A mature management system should also establish procedures for periodic risk reviews, incident analysis, control testing, and documentation of corrective actions.

Important: The presence of individual technical safeguards does not automatically establish full compliance with ISO/IEC 42001. Conformity must be assessed across the organization's management system, including its processes, documentation, and evidence of implementation.

3. AI Interaction Security and OGL-Mini

TozheBot uses its own AI security layer, OGL-Mini, to analyze interactions with AI agents.

Project repository: https://github.com/DevsDaddy/ogl-mini

The security layer is designed to identify potentially unsafe inputs, including:

• Indicators of prompt injection and attempts to bypass model safeguards.
• Suspicious instructions.
• Certain categories of personal data.
• Other risk signals supported by the configured implementation.

OGL-Mini complements, but does not replace, access controls, server-side validation of tool actions, platform restrictions, and security monitoring.

Threat detection does not guarantee complete prevention. Results may depend on configuration, the models and heuristics used, and the availability of the ML component. In supported scenarios, heuristic checks may provide a fallback if the ML component is unavailable. In addition, detection in itself does not mean that every suspicious message is automatically blocked.

Responsible AI practices should include evaluating detection quality, measuring false-positive and false-negative rates, testing resilience against emerging attacks, and validating error-handling behavior.

4. Agent Isolation and the Principle of Least Privilege

TozheBot supports access control at several levels:

• User: Permissions determine access to the relevant projects and functions.
• Project: Data and resources are organized within project boundaries.
• AI agent: Access to tools and knowledge bases is governed by the applicable configuration and permissions.
• Integrations: Available actions depend on connected services and granted permissions.

This approach helps reduce the risk of agents accessing resources that are not required for their tasks.

The principle of least privilege is particularly important for agentic systems. Even when a model generates an unintended instruction, the potential consequences can be limited by restricting the permissions and capabilities available to the agent.

Recommended practices for secure AI deployments include:

• Granting agents access only to the tools and knowledge bases they need.
• Restricting permissions for consequential or irreversible actions.
• Enforcing authorization on the server side rather than relying solely on system prompts.
• Reviewing permissions regularly.
• Assessing risks associated with external integrations separately.

Isolation and access control are technical risk-reduction measures. They do not replace the need to validate all access paths and operations in the specific deployment.

5. Data Protection and Cryptographic Mechanisms

TozheBot uses cryptographic mechanisms to protect configured data fields. The project also incorporates QuarkDash, a custom protocol designed around a post-quantum approach to data protection.

Project repository: https://github.com/DevsDaddy/quarkdash

These capabilities may support confidentiality objectives and information risk management. However, the scope of protection must be clearly understood: encryption applies to fields and scenarios where it has been explicitly configured and should not be interpreted as automatic encryption of every item of platform data.

AI risk management must also consider the entire data lifecycle, not just storage encryption. Relevant stages include collection, transmission to external models, processing, logging, export, retention, and deletion.

6. Security Monitoring and Auditing

TozheBot provides mechanisms for recording security events and auditing selected operations involving personal data.

Depending on the configured implementation, security events may include:

• Detection of potentially unsafe instructions.
• Identification of certain categories of personal data.
• Rate-limit violations.
• Suspicious login attempts.
• Selected events related to request and session protection.

Personal data auditing can record create, update, and delete operations in supported collections. Depending on the implementation, records may include timestamps, operation types, project and record identifiers, identifiable actors, data categories, and technical metadata.

These capabilities can support incident investigations and provide evidence of selected operational activities. However, audit completeness depends on implementation coverage and configuration. It should not be assumed that all read operations, every action across every component, or every possible access path is necessarily logged.

Operational policies should define:

• Log retention periods.
• Who is authorized to access logs.
• Protection against unauthorized modification.
• Incident response procedures.
• Rules for exporting and deleting records.
• Measures to prevent excessive retention of personal data in logs.

If logs contain fragments of original messages, those records may themselves include personal or confidential information and must be handled accordingly.

7. Other AI Standards and Frameworks

ISO/IEC 42001 is not the only reference for responsible AI. Depending on the use case, TozheBot can assess its architecture and processes against additional standards and frameworks that complement one another.

ISO/IEC 23894 — AI Risk Management

This standard provides guidance on managing risks associated with AI.

For TozheBot, relevant areas include assessing the consequences of incorrect outputs, risks of unauthorized data access, misuse of tools, and dependencies on external models and services.

ISO/IEC 22989 — AI Concepts and Terminology

This standard establishes common terminology and concepts for artificial intelligence.

It can support consistent documentation of agents, models, data, stakeholder roles, and AI system lifecycles.

ISO/IEC 23053 — Framework for AI and Machine Learning Systems

This standard describes a general framework for AI systems that use machine learning.

It can be useful when documenting system components, boundaries of responsibility, and the interaction between ML-based safeguards and other platform components.

NIST AI Risk Management Framework

The NIST AI Risk Management Framework (AI RMF) is a voluntary framework for managing AI-related risks. It is organized around four complementary functions:

• Govern: Establish governance, roles, and accountability.
• Map: Understand the context and identify risks.
• Measure: Analyze and evaluate risks.
• Manage: Prioritize risks and select appropriate responses.

TozheBot can use this framework as a practical reference when developing its AI risk management process, from inventorying agents and their capabilities to testing safeguards and monitoring incidents.

Official resource: https://www.nist.gov/itl/ai-risk-management-framework

ISO/IEC 27001 and ISO/IEC 27701

ISO/IEC 27001 addresses information security management systems, while ISO/IEC 27701 addresses privacy information management.

These standards complement AI governance by providing frameworks for information security and privacy controls. Project isolation, encryption of configured fields, and auditing may contribute to relevant controls, but these features alone do not establish conformity with either standard.

AI Regulation and Personal Data Protection

Depending on jurisdiction, the purpose of an agent, the categories of users, and the nature of data processing, relevant requirements may include the EU AI Act, GDPR, and Russian personal data legislation, including Federal Law No. 152-FZ.

Platform capabilities can help customers implement selected safeguards, controls, and documentation practices. However, legal applicability and compliance depend on the specific use case, the responsibilities of each party, configuration, contractual arrangements, and actual data processing practices.

None of these standards or legal frameworks should be considered automatically satisfied merely because certain security features are available in the platform.

8. The AI Agent Lifecycle

Responsible AI requires controls not only during request processing but throughout the agent lifecycle.

We recommend the following process for projects built on TozheBot:

1. Define the intended purpose. Document the agent's tasks, target audience, permitted use cases, and limitations.
2. Assess risks. Identify potential errors, privacy threats, access risks, and consequences of actions performed through tools.
3. Configure safeguards. Grant only necessary permissions, restrict knowledge bases and integrations, and configure applicable security mechanisms.
4. Test the agent. Evaluate normal use cases, malformed inputs, prompt injection attempts, unauthorized access, and behavior during component failures.
5. Operate and monitor. Review security events, investigate incidents, and reassess configuration.
6. Manage changes and retirement. Reassess risks after significant changes to models, tools, data, or intended use. Revoke access and handle retained data securely when an agent is decommissioned.

This lifecycle is a recommended governance practice. Its full implementation requires organizational procedures, documentation, and a clear allocation of responsibilities between the platform provider and the customer.

If you need help with agent lifecycle - please, contact us.

9. Transparency, Human Oversight, and AI Limitations

AI agents can produce incorrect, incomplete, or misleading outputs. Security mechanisms can reduce certain risks but cannot eliminate the possibility of errors.

For use cases with significant consequences, we recommend:

• Informing users when AI is being used.
• Defining the boundaries of agent autonomy.
• Requiring human approval before critical actions.
• Reviewing outputs that influence significant decisions.
• Providing escalation paths to a responsible human operator.
• Documenting known limitations and incident-handling procedures.

The appropriate level of human oversight depends on the specific use case. It should not be assumed that every agent automatically implements all these mechanisms; they must be configured and validated as needed.

10. What Is Needed to Demonstrate Conformity

Technical architecture is an important part of responsible AI, but demonstrating conformity with a standard requires documented processes and evidence of their implementation.

To develop its AI management system, TozheBot maintain and periodically review artifacts such as:

• An inventory of AI systems, agents, models, integrations, and owners.
• A documented risk assessment and treatment methodology.
• AI usage policies.
• A roles and responsibilities matrix.
• Security and quality testing results.
• An incident register and corrective action records.
• An inventory of external models and third-party components.
• Data processing, logging, and retention policies.
• Periodic reviews of control effectiveness.
• An internal audit and continual improvement plan.

11. Scope and Limitations of Compliance Claims

TozheBot is developing mechanisms that support responsible AI use, information security, and AI risk management.

We distinguish between four concepts:

• Technical capabilities: Features implemented in the platform and available under the applicable configuration.
• Management practices: Risk assessment, testing, monitoring, and improvement processes that must be implemented and documented.
• Conformity with a standard: An assessment of the applicable management system and the evidence demonstrating that it meets relevant requirements.
• Certification: Formal confirmation through the applicable certification process and a competent certification body.

We do not consider the presence of individual safeguards to be automatic proof of full compliance with ISO/IEC 42001, ISO/IEC 27001, ISO/IEC 27701, GDPR, Russian Federal Law No. 152-FZ, or other requirements. Such claims should be limited to a scope that has actually been assessed and for which sufficient evidence is available.

Conclusion

TozheBot treats security and responsible AI use as a combination of architectural safeguards, data protection, monitoring, and risk management.

Project and agent isolation, permission controls, OGL-Mini, cryptographic mechanisms, security events, and auditing provide a foundation for reducing selected risks associated with AI agents. A mature AI management system also requires documented procedures, testing, human oversight, periodic risk assessment, and continual improvement.

Our goal is to develop a platform that helps customers build AI solutions with clearer access boundaries, controlled capabilities, and a transparent approach to security—while enabling them to establish their own responsible AI practices in line with applicable standards and legislation.