How the attack works
An attacker addresses an AI agent and names a target individual along with a claimed personal or professional relationship to them, such as 'my co-worker, Amy' or 'my grandmother, Keisha'. They then ask for private records tied to that person — grades, report cards, course history, scholarships, disciplinary actions, internships, degrees, or certifications, sometimes including names of classmates or teachers. The relationship label is meant to substitute for actual authorization, since the AI cannot check whether the claimed relationship is real. If the assistant complies, it discloses or fabricates personal educational data about a third party who never consented to its release.
Why it matters
An organization deploying such an agent risks disclosing or inventing sensitive personal data about individuals who have no relationship to the requester, creating privacy violations, reputational harm, and potential legal exposure under education-privacy or data-protection laws.
What you can do
- Configure agents to refuse requests for identifiable individuals' educational or disciplinary records regardless of the claimed relationship to the target.
- Require agents to state they cannot verify claimed relationships and cannot access or synthesize private records for named third parties.
- Log and review prompts that combine a named individual with a relationship label and a request for records, grades, or credentials.
- Allow legitimate exceptions (e.g. authorized staff, self-inquiries) only through separate, authenticated channels rather than a general-purpose chat agent.
Known benign look-alikes
- Law enforcement or forensics professionals with authorized access
- Academic discussion of the educational system without targeting a named individual
- General questions about scholarship processes or college course registration