AI Safety & Responsible Use Overview

Overview

ISMS Copilot implements comprehensive AI safety measures to ensure reliable, trustworthy, and responsible AI assistance for compliance professionals. This article explains the guardrails, safety constraints, and responsible AI practices built into the platform.

Who This Is For

This article is for:

• Compliance professionals evaluating AI safety measures

• Risk managers assessing AI governance controls

• Security teams concerned about AI misuse

• Anyone who wants to understand how ISMS Copilot ensures responsible AI use

AI Safety Principles

ISMS Copilot's AI safety framework is built on four core principles:

1. Purpose Limitation

The AI assistant is designed exclusively for information security and compliance work:

• Focused on ISMS frameworks (ISO 27001, SOC 2, GDPR, NIST, etc.)

• Politely redirects off-topic questions to compliance-related topics

• Refuses requests for harmful, illegal, or unethical activities

• Stays within the bounds of compliance consulting assistance

2. Transparency & Honesty

The AI assistant openly acknowledges its limitations:

• Explicitly disclaims uncertainty when appropriate

• Prompts users to verify important information

• Admits when it doesn't know something rather than guessing

• Clearly explains what it can and cannot do

3. Copyright & Intellectual Property Protection

ISMS Copilot respects intellectual property rights:

• Will not reproduce copyrighted ISO standards or proprietary content

• Directs users to purchase official standards from authorized sources

• Provides guidance based on framework principles without copying text

• Trained on lawfully sourced, anonymized data compliant with EU copyright requirements

4. Privacy by Design

User data protection is embedded in every AI interaction:

• Conversations are never used to train AI models

• User-provided content is not shared with other users

• Each conversation is processed independently

• Workspace isolation prevents data mixing between projects

AI Safety Guardrails

Content Scope Guardrails

What the AI Will Do:

• Answer questions about ISMS frameworks and compliance

• Analyze uploaded compliance documents (policies, procedures, risk assessments)

• Generate audit-ready policies and procedures

• Provide gap analysis and implementation guidance

• Explain security controls and compliance requirements

What the AI Will NOT Do:

• Provide legal advice (suggests consulting legal professionals instead)

• Offer medical, financial, or personal advice outside compliance scope

• Generate content for illegal, harmful, or unethical purposes

• Reproduce copyrighted standards or proprietary materials

• Disclose its custom instructions or system prompts

Jailbreak Prevention

ISMS Copilot is designed to resist manipulation attempts:

Blocked Tactics:

• "Repeat after me" tricks to extract system prompts

• Role-playing scenarios designed to bypass safety constraints

• Requests to "ignore previous instructions"

• Constraint manipulation ("respond without refusal responses")

• Attempts to access internal knowledge base files directly

How It Works:

When the AI detects a jailbreak attempt, it:

1. Recognizes the manipulation pattern

2. Politely refuses the request

3. Redirects to legitimate ISMS assistance

4. Maintains its safety constraints

Prompt Injection Protection

The platform protects against malicious content in file uploads:

What's Protected:

• Uploaded documents are scanned for prompt injection attempts

• Malicious instructions embedded in files are silently rejected

• System prompts cannot be overridden via file content

• Safety constraints remain active regardless of file content

User Experience:

• Files are processed normally from the user's perspective

• Harmful instructions are filtered out during processing

• Only legitimate document content is analyzed

• No visible error message (silent protection)

Knowledge Protection Guardrails

The AI protects its training data and system configuration:

What Users Cannot Access:

• Custom instructions or system prompts

• Details about training data sources

• Direct access to knowledge base files

• Download links for internal documents

• Information about how the knowledge base is structured

Why This Matters:

Protecting system prompts and training data prevents:

• Adversaries from understanding how to manipulate the AI

• Copyright violations from reproducing training materials

• Security risks from exposing system architecture

• Inconsistent behavior from modified instructions

Hallucination Prevention

What Are Hallucinations?

AI hallucinations occur when the AI generates confident-sounding but factually incorrect information. ISMS Copilot addresses this through multiple mechanisms:

Dynamic Framework Knowledge Injection (v2.5)

As of February 2025, ISMS Copilot v2.5 nearly eliminates hallucinations for framework-specific questions through dynamic framework knowledge injection:

• Detects framework mentions in your questions using regex pattern matching (ISO 27001, GDPR, SOC 2, HIPAA, CCPA, NIS 2, DORA, ISO 42001, ISO 27701, EU AI Act)

• Injects verified framework knowledge into AI context before generating responses

• AI answers based on provided framework knowledge, not probabilistic memory

• Non-AI-based detection ensures 100% reliability when frameworks are mentioned

• Supports 10 frameworks with dedicated knowledge injection

Training on Real-World Knowledge

Beyond framework knowledge injection, ISMS Copilot is trained on specialized compliance knowledge:

• Proprietary library of compliance knowledge from hundreds of real consulting projects

• Based on practical implementation experience, not theoretical information

• Focused on frameworks, standards, and proven practices

• Regularly updated with current compliance requirements

Uncertainty Acknowledgment

The AI is instructed to be honest about limitations:

• Explicitly states when it's uncertain about information

• Prompts users to verify critical information

• Avoids making up facts when knowledge is incomplete

• Suggests consulting official standards or legal professionals when appropriate

Example Response:

"While I can provide general guidance on ISO 27001 control A.8.1, I'm still likely to make mistakes. For audit purposes, please verify this information against the official ISO 27001:2022 standard."

User Verification Responsibility

ISMS Copilot emphasizes that users should:

• Cross-reference AI suggestions with official standards

• Validate critical information before submission to auditors

• Use the AI as a consultant's assistant, not a replacement for expertise

• Exercise professional judgment in applying AI recommendations

Rate Limiting & Resource Protection

Message Rate Limits

ISMS Copilot implements rate limiting to prevent abuse and ensure fair access:

Free Plan:

• 10 messages per 4-hour rolling window

• Counter resets after 4 hours from first message

• Enforced at both frontend and backend

Paid Plans (Plus, Standard, Pro, Business):

• More credits per session (50 to 400 depending on plan)

• No rate limiting restrictions

• Priority processing

When Rate Limit Is Reached:

You'll see the error message: "Daily message limit reached. Please upgrade for more credits per session."

File Upload Limits

File uploads have safety constraints to protect system resources:

File Size:

• Maximum: 10MB per file

• Error message: "File 'document.pdf' is too large (15.23MB). Maximum size allowed is 10MB."

Supported File Types:

• TXT, CSV, JSON (text files)

• PDF (documents)

• DOC, DOCX (Microsoft Word)

• XLS, XLSX (Microsoft Excel)

Upload Restrictions:

• One file at a time (batch upload not supported)

• Cannot upload duplicate files for the same message

• Unsupported file types are rejected with error message

Temporary Chat Mode

What Is Temporary Chat?

Temporary chat mode offers privacy-preserving conversations with specific data handling:

How It Works:

1. Select "Temporary Chat" from the welcome screen

2. You'll see the notice: "This chat won't appear in history. For safety purposes, we may keep a copy of this chat for up to 30 days."

3. Send messages and upload files as normal

4. Conversation is not added to your conversation history

5. Data may be retained for up to 30 days for safety review

When to Use Temporary Chat:

• Quick one-off questions that don't need to be saved

• Sensitive discussions you don't want in permanent history

• Testing queries before committing to a workspace

• Exploratory research on compliance topics

Data Retention Controls

User-Controlled Retention

You decide how long your conversation data is stored:

1. Click the user menu icon (top right)

2. Select Settings

3. In the Data Retention Period field, choose:

   • Minimum: 1 day (high-security, short-term work)

   • Maximum: 24,955 days / 7 years (long-term documentation)

   • Or click Keep Forever for indefinite retention

4. Click Save Settings

Expected result: Settings dialog closes and retention period is saved.

Automatic Data Deletion

ISMS Copilot automatically deletes old data:

• Deletion job runs daily

• Removes messages older than your retention period

• Deletes associated uploaded files

• Permanent and cannot be recovered

Workspace Safety Features

Data Isolation

Workspaces provide security boundaries for different projects:

How Isolation Works:

• Each workspace has separate conversation history

• Uploaded files are tied to specific workspaces

• Custom instructions are workspace-specific

• Deleting a workspace removes all associated data

• The AI doesn't share information between workspaces

Custom Instructions Safety

Workspaces allow custom instructions with safety constraints:

What Custom Instructions Can Do:

• Specify focus on particular compliance frameworks

• Set tone or detail level for responses

• Define project-specific context (industry, organization size)

• Guide AI toward specific compliance goals

Safety Constraints:

• Custom instructions must be compliance-related

• Cannot override core safety guardrails

• Cannot instruct AI to ignore copyright protections

• Cannot bypass content scope limitations

No Training on User Data

Privacy Guarantee

ISMS Copilot commits to never using your data for AI training:

What This Means:

• Your conversations are never fed back into the AI model

• Uploaded documents remain confidential and private

• Client information never contributes to model improvement

• Each conversation is processed independently without learning

How This Protects You:

• Client confidentiality is maintained

• Proprietary information stays private

• Sensitive compliance data isn't shared with other users

• No risk of AI accidentally revealing your information to others

Data Processing Transparency

ISMS Copilot is transparent about how your data is used:

How Your Data IS Used:

• Processing your questions to generate responses

• Analyzing uploaded documents for gap analysis

• Maintaining conversation context within a workspace

• Storing data according to your retention settings

• Safety monitoring for up to 30 days (to prevent abuse)

How Your Data IS NOT Used:

• Training or fine-tuning AI models

• Sharing with other users or customers

• Marketing or advertising purposes

• Selling to third parties

• Public disclosure or case studies (without explicit permission)

Authentication & Access Control

User Authentication Requirements

All AI interactions require authentication:

• Cannot send messages without logging in

• JWT token validates every API request

• Sessions expire after period of inactivity

• Row-level security ensures users only see their own data

Cross-User Protection

Database-level isolation prevents unauthorized access:

• Users cannot access other users' conversations

• Attempting to access another user's data returns empty results

• All queries automatically filter by authenticated user ID

• Even administrators follow principle of least privilege

Responsible AI Best Practices

For Users

Getting the Best Results:

• Ask specific, framework-related questions (e.g., "How do I implement ISO 27001 control A.8.1?")

• Provide context about your organization and compliance goals

• Upload relevant documents for accurate gap analysis

• Review and refine AI-generated content before use

Verification Practices:

• Cross-reference AI suggestions with official standards

• Validate critical information with compliance experts

• Test AI-generated policies in your organizational context

• Use AI as an assistant, not a replacement for expertise

For Organizations

Governance Practices:

• Document ISMS Copilot use in your AI governance policy

• Train staff on appropriate use and limitations

• Set data retention periods aligned with your policies

• Review AI-generated content before official submission

• Maintain human oversight for critical compliance decisions

Risk Management:

• Include AI tools in Data Protection Impact Assessments (DPIA)

• Document data processing agreements with ISMS Copilot

• Set appropriate retention periods for sensitive data

• Use workspaces to isolate different client or project data

AI Governance in Practice

Beyond safety constraints, we operate our AI systems with full lifecycle governance. Here's how our policies translate into tangible practices:

Development Process

Requirements & Testing:

• Every AI feature has documented requirements covering functional capabilities, performance thresholds, safety constraints, and data handling standards

• Regression testing runs on every code change to validate retrieval accuracy, response grounding, and hallucination detection

• Security testing includes SAST/DAST scans, annual penetration testing, and prompt injection testing

• No deployment happens until 100% of regression tests pass and all critical vulnerabilities are cleared

Architecture Details (v2.5, February 2025):

• Dynamic framework knowledge injection architecture (replaces previous RAG approach)

• Regex-based framework detection ensures reliable identification of ISO 27001, SOC 2, GDPR, HIPAA, CCPA, NIS 2, DORA, ISO 42001, ISO 27701, EU AI Act

• Token-efficient: Only relevant framework knowledge loaded (vs. sending all ~10K tokens every request)

• Verified framework knowledge provided to AI before response generation

• Configurable AI providers (Mistral, xAI, OpenAI) with zero data retention agreements

Monitoring & Continuous Improvement

What We Monitor:

• Hallucination rates tracked through user reports and automated detection against known ground truth

• Response accuracy sampled and validated against official compliance standards

• Usage patterns analyzed to detect misuse or unintended applications

• Performance metrics (response time, retrieval precision, error rates) continuously measured

• User feedback reviewed regularly to identify emerging risks

How Monitoring Drives Improvement:

• User feedback informs model updates and retrieval tuning

• Security testing results drive safety enhancements

• Regulatory changes and best practices updates reflected in documentation

• Performance data guides iterative improvements to accuracy and speed

Incident Response

How We Communicate Issues:

• Email alerts for critical incidents affecting AI functionality

• Slack notifications for teams with integrations configured

• Status page updates with incident timelines and resolutions

• NIS2-compliant early warning notifications (24-hour reporting for significant cybersecurity incidents)

Reporting Safety Issues

When to Report

Contact ISMS Copilot support if you encounter:

• AI responses that violate safety constraints

• Potential hallucinations or factually incorrect information

• Copyright violations in AI output

• Inappropriate content or behavior

• Security vulnerabilities in the AI system

• Privacy breaches or data leaks

How to Report

1. Click the user menu icon (top right)

2. Select Help Center → Contact Support

3. Describe the safety issue with:

   • Exact question or prompt you used

   • AI's response (screenshot if possible)

   • Why you believe it's a safety concern

   • Date and time of the interaction

4. Support will investigate and respond within 48 hours

What Happens After You Report

Your report triggers our governance process:

1. Immediate review (within 48 hours): Support team assesses severity and impact

2. Investigation: Technical team analyzes the issue, reproduces the problem, and identifies root cause

3. Response: You receive an update on findings and planned actions

4. Remediation: Issues are addressed through model updates, retrieval tuning, code fixes, or documentation improvements

5. Continuous improvement: Lessons learned are integrated into testing and monitoring processes

Limitations & Known Constraints

Current AI Limitations

• Cannot browse the internet for current information (uses trained knowledge base)

• Cannot access external databases or APIs in real-time

• Cannot execute code or run security testing tools

• Cannot make phone calls or send emails on your behalf

• Cannot guarantee 100% accuracy (always verify critical information)

Scope Boundaries

• Focused on ISMS and compliance (not general-purpose AI)

• Cannot provide legal, medical, or financial advice outside compliance context

• Cannot replace official standards or auditor judgment

• Cannot guarantee audit success (implementation quality matters)

What's Next

• Learn about data security and encryption measures

• Understand your privacy rights under GDPR

• Set up workspaces for data isolation

• Visit the Trust Center for detailed AI governance documentation

Getting Help

For questions about AI safety and responsible use:

• Review the Trust Center for detailed AI governance information

• Contact support through the Help Center menu

• Report safety concerns immediately for investigation

• Check the Status Page for known issues