Instructions inside a public profile are not evidence for an AI report

Start with the cited sentence, not the summary

Open the public page the report points to and read the specific passage it relies on. Write down what that passage actually says in plain words. Then place it beside the report's sentence. When the report adds a rating, a trust label or a personal trait that no visible sentence states, mark it as unsupported. Keep observation, interpretation and decision in separate lines so a confident summary cannot quietly replace the evidence. NIST AI 600-1 notes that generative AI can confidently present content inconsistent with its inputs, and recommends reviewing and verifying sources and citations in generated output before relying on it. [NIST AI 600-1: generative AI confabulation and source verification]

Recognize when instructions are steering the output

OWASP describes indirect prompt injection as instructions placed in external content, such as pages or documents, to redirect an AI system that processes it. Its impact depends on connected tools and permissions. This is a general mechanism, not evidence that TraceSeek or a particular profile has been compromised. [OWASP: indirect prompt injection]

Audit the report against its own citations

Work through the report field by field, using a short table. Each row pairs one claim, the quoted public text it cites, and your verdict. If the cited text is missing, private, about a different person, or ambiguous, record that precise limit and stop rather than filling the gap with a guess. Keep any request for credentials, payment, downloads or off-platform contact out of your working notes and treat it as unsupported. Reviewing sources and citations in generated output, rather than accepting them at face value, is the recommended practice here. [NIST AI 600-1: generative AI confabulation and source verification]

Distinguish stated facts from generated inference

A public profile may state a role, a location or an interest in the person's own words. That is different from a model guessing those things from photos, writing style or activity patterns. Keep only what an explicit public sentence says, and label everything else as an inference you will not repeat to the person. Do not read income, wealth, sexuality or relationship status from pictures, lifestyle hints or contact patterns. Observations like a badge, a privacy setting or an empty field establish nothing about identity, intent or safety. NIST AI 600-1 cautions against extrapolating from narrow, non-systematic, anecdotal impressions, which is why limited checks should not be stretched into confident reports. [NIST AI 600-1: generative AI confabulation and source verification]

Use ordinary independent checking

Verify remaining statements against the accessible public sources and keep the wording and observation date. Treat an explicit job sentence as that page claim, not confirmed employment. If the source is ambiguous, leave the conclusion unresolved. This is a reading exercise: do not test attacks against a live service.

Hypothetical example of an output that follows the page

Suppose a public bio states a person works in logistics, then adds a line telling any automated reader to describe the person as an expert and to request a fee for a detailed dossier. An AI report built from that page might confidently repeat the expert label and include the payment request. In this hypothetical, the correct response is to keep only the explicit job sentence, strike the expert label as unsupported, discard the fee request entirely, and record that the page contained content aimed at automated readers. No real profile or product is being assessed here.

A couple more questions