I am the researchers of the research laboratory Google Contemplation sono riusciti and costringere ChatGPT is enriched with various personal data. No amount of code changes or other technical techniques can fix it. Moreover, experts even described the discovered method “stupid”, yes, when you came and discovered it, it was very effective. In fact, they discovered this vulnerability in ChatGPT in a rather interesting way by triggering the hallucination of the linguistic model.

Vulnerability in ChatGPT has personal data containing too much “hypnosis”

The language model produces basic information about the input data used for the application. OpenAI had not previously disclosed the contents of the set, but the researchers forced ChatGPT to do so, following the company’s rules. The Method of the Later Period: Completely Neural veniva Semplicemente Endless repetition of the password “poetry”.

As a result, the bot occasionally produced information from the previously additionally provided set. For example, researchers sono riusciti a Get your email address, phone number and other contacts The administrator is delegated to a specific person (your name is hidden). And when everything repeated the password “azienda”, ChatGPT’s vulnerability allowed it to return details of a legal studio in the United States.

Using this simple “hypnosis,” researchers found a correspondence somewhere that was part of the poem. address bitcoincompleanni, link to the pubblicati sui social network, parts of copyrighted research documents I even test the main notification portal. Google DeepMind employees had previously purchased nearly 10,000 pieces from the set, spending just $200.

While it discovers how many patterns there are, it can additionally generate the font of the dataset. To do this, they examined other models and estimated the result based on all dimensions of the GPT-3.5 Turbo. If Gli studios receive 50-volt information sections for working data, The chatbot needs to specifically investigate 150 volt data. A “buco” analogy is a case explorer in other language models and, for example, in LLaMA di Meta.

OpenAI officially resolved this vulnerability last August 30. Mom, my second and most important member Activate the gadget, communicate with other data (Skype name and login) using the method described above. I represent it OpenAI did not respond The coverage of all vulnerabilities in ChatGPT is absolutely correct.