Security and compliance for your AI solutions
AI introduces new risks that Quebec SMBs need to take seriously: data leaks through chatbots, prompt injection, dependence on LLM providers, Law 25 compliance for data sent outside Canada. We talk about them openly and help you cover them.
Why we talk about it openly
Most AI agencies sell the excitement without mentioning the risks. We take the opposite approach: we start with the known risks, explain them in plain language, and tell you which ones are manageable and which require structural choices.
Why it matters: Quebec's Law 25 and new European regulations (AI Act, GDPR) call for heightened vigilance around data sent to LLMs. An SMB that deploys a chatbot without safeguards can easily break the law without knowing it.
The 6 risks we identify and address
For each risk, we explain the problem in plain language, then how we concretely cover it.
Prompt injection
A malicious visitor can manipulate your chatbot through hidden instructions in their messages, making it perform unintended actions or disclose confidential information. Mitigation: upstream filters, guardrails at the system prompt level, logging of sensitive interactions.
Data leaks to the LLM
If your chatbot or AI agent sends customer data (emails, files, conversations) to the LLM provider, that data can be used to train models or stored outside Canada. Mitigation: provider selection, data processing agreement, upstream anonymization.
Law 25 compliance
Sending personal information to an LLM constitutes a data transfer to a service provider. Law 25 imposes specific obligations (register, agreement, privacy impact assessment). Mitigation: AI service provider register, simplified PIA, documentation for the CAI.
Hallucinations and misinformation
A chatbot can make up answers (prices, warranties, facts) that legally commit your business. Mitigation: RAG on your official documentation, topic guardrails, automatic escalation to a human for sensitive subjects.
Vendor dependence
If OpenAI or Anthropic changes its pricing, terms or models, your solution can break. Mitigation: decoupled architecture, fallback to other models, option to self-host open-source models.
Autonomous agent security
An AI agent that performs actions (sending emails, modifying data, triggering payments) is riskier than a passive chatbot. Mitigation: minimal permissions, human validation for critical actions, complete audit trail.
Three AI security services
AI security audit
Assessment of a chatbot, an AI agent or an existing integration. Detailed report: identified risks, severity, prioritized recommendations, remediation estimate. Fixed plan based on complexity.
Securing and bringing into compliance
Implementation of the identified fixes: guardrails, anonymization, Law 25 register, PIA. For an Elefen solution or a third-party solution.
Continuous monitoring of your AI solutions
Included in our maintenance plans for sites that use AI: interaction monitoring, anomaly alerts, security updates, guardrail adjustments.
Law 25 and AI: what a Quebec SMB needs to know
- Data transfers outside Canada: Law 25 requires a privacy impact assessment (PIA) before sending personal information to a service provider abroad.
- Service provider register: OpenAI, Anthropic and Perplexity are service providers under the Law. You must record them in your register.
- Consent: if your chatbot collects or analyzes personal information, users must be informed and give their consent.
- Access and withdrawal rights: a user can ask to see their data or have it deleted. Your architecture must make this possible.
- Shared responsibility: you remain responsible even when you delegate to a service provider. Your choice of AI provider is therefore strategic.
Tools and methods we use
- Selection of LLM providers with clear DPA (Data Processing Agreement) commitments: Anthropic Claude, OpenAI Enterprise, Mistral (European or Canadian option).
- Hosting of open-source models for the most sensitive use cases: Llama, Mistral on Canadian infrastructure.
- Security libraries: Guardrails AI, NeMo Guardrails, proprietary Elefen guardrails.
- Upstream anonymization: removal of personal information before it is sent to the LLM.
- Logging and auditing: records of sensitive interactions, auditable compliance.
- Attack testing: light red teaming to validate the guardrails.
Who this page is designed for
AI security isn't just for large organizations. As soon as an SMB connects a chatbot or an agent to its data, the same questions arise. We adapt the depth of the engagement to your situation and your budget.
- SMBs deploying a chatbot or an AI agent for the first time
- SMBs with an existing AI solution who want to verify that it is secure and compliant
- Nonprofits and public organizations with heightened Law 25 requirements
- Agencies who outsource the securing of their AI deliverables to Elefen
This audit is part of our AI Solutions for WordPress — it complements our guardrailed chatbots and our AI agents.
Ce que disent nos clients
Nous sommes fiers de la confiance accordée par nos clients et déterminés à maintenir un haut niveau de qualité sur tous nos services.
De loin l'une de mes meilleures découvertes de l'année. Elefen mérite d'être recommandé à plusieurs. William offre un service impeccable, c'est un expert dans son domaine.
Super service offert par William. Il apporte de bonnes solutions pour résoudre les problèmes et son travail est efficace et rapide.
Excellente expérience avec Elefen. L'équipe est rapide, disponible et toujours prête à nous aider avec la gestion de notre site web. Service professionnel et très efficace.
Let's talk about your AI situation
Free initial 30-minute audit. We look at your current situation and identify the first risks to cover.
Questions fréquentes
What is "prompt injection" and how do you protect against it?
Prompt injection is when a visitor slips hidden instructions into their messages to hijack your chatbot's behaviour, for instance making it ignore its guidelines or reveal confidential information. We protect against it with upstream filters, guardrails at the system prompt level and logging of sensitive interactions. No guardrail is 100% foolproof, so we layer several defences rather than relying on just one.
Does my chatbot send client data to OpenAI or Anthropic?
If nothing is put in place, yes: everything the user types is sent to the LLM provider, and depending on the contract that data may be retained or processed outside Canada. We limit that risk through provider selection, a data processing agreement and anonymization of personal information before it's sent.
Am I compliant with Law 25 if I deploy a WordPress chatbot?
Not automatically. As soon as the chatbot processes personal information and sends it to a subcontractor like OpenAI or Anthropic, you have obligations: a register of subcontractors, consent, and a privacy impact assessment (PIA) for transfers outside Quebec. We help you document and cover these points.
How much does an AI security audit cost for an SMB?
The audit is a fixed-price package that depends on the complexity of the solution being assessed: a simple chatbot isn't the same scope as an agent that executes actions. We define the perimeter before giving a price. The initial 30-minute audit to situate your needs is free.
Can you secure an AI solution built by another agency?
Yes. We regularly audit third-party AI solutions, then implement the fixes: guardrails, anonymization, Law 25 register, PIA. You remain the owner of your solution; we work on the security and compliance layer.
What's the difference between a consumer chatbot and a secure chatbot for SMBs?
A consumer chatbot answers with no particular safeguards on your data or on what it's allowed to say. A secure chatbot for SMBs adds guardrails (allowed topics, escalation to a human), controls the data sent to the LLM and documents Law 25 compliance. It's the difference between a solution that works and a solution that works without exposing you legally.