keepcalculator

06 Numbers and Powers

AI Cybersecurity Incidents: What to Tell Users

Card No. 06·81316 · Updated · Standard formula, shown below

When an AI system your business uses is involved in a cybersecurity incident, affected users need a clear, honest, and practical notification that explains what happened, what data was exposed, and what they should do next. This article outlines the key elements of an effective user notification, including legal considerations, communication channels, and common pitfalls to avoid.

Background: Why AI Incidents Differ

AI systems introduce unique notification challenges. Unlike a traditional database breach where you know exactly which records were accessed, AI incidents often involve models that were trained on user data, or that generate outputs based on prompts. Users may not have a clear mental model of what “AI involved” means. For example, if a chatbot leaks another user’s private messages, the affected users are both the person whose data was leaked and the person who saw it. If an AI model is stolen, the training data – which may include personal information – is at risk, but you might not know exactly which pieces were exposed.

Key Elements of a User Notification

When you notify affected users, include these components:

  • What happened – a plain-language description of the incident, avoiding jargon like “model inversion” unless you explain it.
  • What data was involved – be specific if possible: names, emails, payment info, chat logs, etc. If you don’t know, say so.
  • What you are doing – steps taken to contain the incident, such as taking the system offline, revoking API keys, or patching vulnerabilities.
  • What the user should do – concrete actions: change passwords, enable two-factor authentication, monitor accounts, or be wary of phishing.
  • Contact information – a dedicated email, phone line, or web form for questions.
  • Regulatory references – if applicable, mention that you are reporting to authorities (e.g., GDPR, HIPAA, state breach laws) without legal advice.

How to Choose the Right Channel and Timing

Notification can be delivered via email, in-app alerts, SMS, or a public disclosure. Choose based on urgency and the user’s likely location. For high-sensitivity data, email plus in-app notification is common. For very urgent situations (e.g., exposed passwords), SMS can be justified. Timing matters: many regulations require notification “without undue delay” – often 72 hours for GDPR. Do not wait for full forensic analysis if you have enough to inform users of immediate risks. However, avoid sending incomplete or misleading messages that you must later correct.

Common Problems and How to Avoid Them

  • Overly technical language – users don’t care about “adversarial perturbations”; they care about whether their data is safe.
  • Vague statements – “We take your privacy seriously” is not helpful. Be specific about what you know and what you don’t.
  • Delaying notification – waiting weeks to “investigate” can erode trust and violate laws.
  • Not providing next steps – telling users to “stay tuned” without actions leaves them helpless.
  • Forgetting non-technical users – some users may not understand what “AI” means; explain in simple terms.

Comparing AI Incidents to Traditional Breaches

Traditional breaches typically involve stolen credentials or SQL injection, where the data is copied. AI incidents can be more subtle: an AI model might memorize training data and regurgitate it in response to a prompt, or an attacker might use prompt injection to extract sensitive information. The notification should reflect the nature of the exposure. For a traditional breach, you might say “credit card numbers were stolen.” For an AI incident, you might say “the chatbot may have displayed your conversation to another user.” The latter is harder to quantify, but honesty is essential.

Safety and Legal Considerations

Do not include specific security vulnerabilities in a public notification, as that could aid attackers. Instead, say “we have patched the vulnerability.” Also, be careful not to overpromise – do not say “no harm will occur” if you cannot guarantee it. If you are subject to regulations, consult legal counsel before sending. For example, GDPR requires notifying the supervisory authority within 72 hours, and affected individuals “without undue delay.” In the US, state laws vary; the California Consumer Privacy Act (CCPA) has its own requirements. Always err on the side of transparency, but avoid speculative statements.

Frequently Asked Questions

Should I mention that AI was involved?

Yes, if it is relevant to how the incident occurred. Users have a right to know if a novel technology caused the issue, but explain it in plain terms. For example, “Our AI-powered chat assistant inadvertently displayed another user’s data due to a software bug” is clearer than “A model hallucination occurred.”

What if I don’t know exactly what data was exposed?

Say that you are still investigating, but provide a worst-case scenario. For instance, “We believe your email address may have been exposed, but we are confirming whether other data was affected.” This gives users enough to take protective actions without waiting for a final report.

How long should I wait before notifying?

As soon as you have a reasonable understanding of the scope and impact. For many regulations, 72 hours is the maximum, but if you can notify sooner, do so. If you need more time to assess, send an initial alert with limited details and a follow-up later.

Can I notify only some users?

Only if you are certain that other users are not affected. If there is any doubt, notify a broader group with a message that says “we are notifying you because your data may have been involved.” It is better to over-notify than to miss someone.

Should I offer credit monitoring or other free services?

If financial data or Social Security numbers were exposed, offering credit monitoring is a common practice, but it is not mandatory. If you do offer it, be clear about what it covers and how to sign up. For AI incidents involving chat logs, credit monitoring may not be relevant; instead, advise users to review their conversations for sensitive information that might have been leaked.

What if the incident was caused by a third-party AI vendor?

You are still responsible for notifying your users. Coordinate with the vendor to get accurate details, but do not shift blame. Your notification should focus on what you are doing to protect users, not on internal finger-pointing.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.

More cards in this drawer

The whole Numbers and Powers drawer →