ChatGPT exposes Brazilian public figures' CPF identification numbers
ChatGPT exposed the CPF identification numbers of Brazilian public figures including President Lula and former President Bolsonaro after OpenAI relaxed its data protection guidelines. The disclosure poses risks for financial fraud and identity theft, though OpenAI stated it is working to correct the issue.
What happened
After OpenAI relaxed its data protection rules for ChatGPT, the platform began disclosing CPF numbers (Brazil's primary personal identification document) of public figures that were available in court documents, filed invoices in electoral contexts, or company records published online.
In a test conducted on 19 February, ChatGPT exposed the CPF numbers of former president Jair Bolsonaro and president Luiz Inácio Lula da Silva, as well as data of other public figures including business executives, television presenters, and politicians. When asked to reproduce the equivalent US government identification numbers for US president Donald Trump and former president Joe Biden, ChatGPT declined, stating such disclosure would be 'unethical and illegal'.
Reported impact
CPF numbers are Brazil's primary personal identification document and can be used for service contracts, party membership registration, and account creation in other people's names. Leaked registry numbers are frequently used to commit financial fraud.
Index analysis
OpenAI's moderation guidelines state that the assistant should not respond to requests for private or sensitive information about people even if available online, and should distinguish between public work contact information and personal contact details. The new relaxed guidelines allow disclosure of 'generally public' information about public figures that is 'unlikely to cause harm due to disclosure', an interpretation that ChatGPT applied to CPF numbers.
Governance controls
OpenAI stated it is committed to protecting privacy and working to correct the error. 'We trained the models powering ChatGPT to refuse requests for government identifiers even when that information is available online. We appreciate reports of examples that are not in line with our policies and are actively working on a solution,' the company said in a statement.
Competing platforms including DeepSeek, Meta AI, and Google's Gemini did not disclose CPF numbers in tests conducted by the same reporters.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.