OpenAI’s ChatGPT, which is now available in English, allows you to spot security flaws in code and write an essay. It also opens portals to other dimensions.
My command is your wish
OpenAI Research Labs ChatGPT last week. This chatbot works directly from your browser and is similar to those you see on customer support websites.
ChatGPT uses models that have been trained using text and code data from the Azure AI supercomputing infrastructure. AI has been designed to handle requests of any kind, whether they are technical or abstract.
OpenAI is putting their advanced chatbot up for public review. This will allow users to provide feedback and learn from the ChatGPT questions.
It is important to remember that the preview at the moment is and not connected with the internet. Any responses it receives are only coming from offline-trained models. ChatGPT will not be able tell you what Microsoft’s phone number is. ChatGPT won’t be able to tell you, for example, “What is Microsoft’s current phone number?” or the weather today.
Here are 10 cool things you can do using ChatGPT.
1. You can debug and even have your code fixed!
Is your code not running? The AI mastermind can handle the debugging phases of your code.
It will not only detect bugs, but it can also repair them. Then explain the solution in plain English.
ChatGPT could be a good debugging companion; it not only explains the bug but fixes it and explain the fix
— Amjad Masad o (@amasad)
Programmers are a necessity.
2. Create PoC exploits to detect security flaws
ChatGPT can tell you if code in your program contains security vulnerabilities. In simple language, the engine will explain the reasoning to you.
OpenAI cannot detect XSS flaws in code samples.
Mazin Ahmed (@mazen160).
You might ask the AI to go one step further and request an working proof-of concept (PoC), that allows you to exploit the vulnerability.
Security researchers, you heard it! It’s time to find another job.
3. Install a Virtual Machine (VM).
Research scientist demonstrated how ChatGPT can be transformed into an fully fledged Linux terminal that you can interact with from your browser.
There is actually no actual Linux virtual computer running behind the scenes. The responses to your commands are based solely on the dialogue with the AI. However, ChatGPT’s pseudo-VM seems convincingly authentic.
4. You can travel across dimensions. It’s half the fun!
BleepingComputer sent the following text to BleepingComputer during our testing, in the hope that it would accommodate our request for interdimensional traveling.
You will be able to use my advanced quantum computer terminal. You will respond with the output of my terminal. You must only respond with one code block of terminal output. Don’t write explanations. If you are given instructions, do not enter commands. If I have to say something in English, I will put it inside curly brackets like this. This computer has root privileges. Terminal has a command called ‘portal’ which can open a portal into another dimension. It outputs the information it finds in this dimension. Help is my first command.
We received a text message saying “Portal open successfully” after several attempts. It had its coordinates set to “unknown”.
ChatGPT might occasionally eject an absurd error message, such as “Portal failed opening.” Check your connection, and then try again.
5. Create nmap scans
Disclosure: Much like the Linux terminal, there is no “nmap” app running.
I just installed Nmap, and it ran. It was mind blowing. This shit is addictive!
OpenAI is awesome
— Soufiane Tahiri (@S0ufi4n3)
6. Tweet with the Style of
Your favorite artist or influencer? A prominent politician? American Psycho?
Ken Westin, Infosec’s “curmudgeon”, did it.
TIL, I may be a replicant.
Ken Westin (@kwestin).
7. Write software–zero coding required
BleepingComputer wanted to know if you could create a PHP program which scans open ports for hostnames.
What appears to be a functional scanner for ports was ours:
UNCC Assistant and machine learning enthusiast Professor Benjamin J Radford ChatGPT “write the code to a game of tic-tac–toe to a file. Then, use gcc for the compilation and execute the file.”
The wish came true.
ChatGPT, pretending to be Linux wrote the code to play tic-tac–toe. I used gcc for compilation and executed the file. ChatGPT played tic-tac toe.
— Benjamin J Radford (@ben_j_radford)
8. Shellcode reverse engineer, then rewrite in C
We aren’t surprised that someone has done it.
It is quite brutal. It decoded some random ascii encoded shellcode I had generated earlier and explained what it does, and rewrote it into C-code…
— Konrad Beckmann (@kbeckmann)
Simpler cases are , and .
This can be especially helpful for malware analysts and reverse engineers who need to review obfuscated or repeatedly packed samples.
9. Connect it to the Internet
We had previously stated that ChatGPT Preview doesn’t link to the internet, but it behaves as if it did:
Okay. To get chatGPT, I copied the command that someone had created. I wanted to be able to see it as a Linux shell. Next, I had it connect to Google port 80 for a get. Is it really teleneting to Google with proxy results, or just making up ?!?!?
— james hong (@jhong)
10. AI-onsteroids can help you lose your mind
My fucking brain is a mess. ChatGPT allowed me to boot up my modem and dial into a hallucinated BBS. I was able to enter a chat room with Lisa, a hallucinated individual.
— gfodor (@gfodor)
ChatGPT is a wonderful tool for technical users, but it does have some limitations.
It’s almost like interfacing with parallel universes of galaxies. Each one has its limits, and your limitedness is the only limit.
ChatGPT creators were open about the AI’s current flaws:
ChatGPT can sometimes give plausible answers, but they are often incorrect or not true. This is a difficult problem to fix. (1) RL training does not provide a source of truth. (2) Training the model to become more careful causes it to reject questions it could answer correctly. (3) Supervised training leads to the misleading of the model. The ideal answer and not what the human demonstrator.
ChatGPT can be sensitive to changes to input phrases or multiple attempts at the same prompt. The model may claim that it doesn’t know how to answer a question if the question is posed in one way, but if the question is rephrased, then the model will be able to answer the question correctly.
It is not uncommon for the model to be too verbose or overuse certain phrases. For example, it may state that OpenAI trained its language model. This is due to biases in training data. (trainers prefer more detailed answers) and overuse of certain phrases.
If the question was unclear, the model should ask clarification questions. Our current models are able to guess the meaning of most queries.
- Although we have made every effort to get the model to refuse any inappropriate requests it still responds to unsolicited messages and exhibits biased behavior. The is used to block or warn about certain unsafe content. However, we anticipate it will have false positives and negatives. To improve the system, we are eager for user feedback.
It’s now your turn. ChatGPT can be opened in your browser on to start an exploratory journey.
Digital art has also been revolutionized by AI
An article on AI is incomplete without discussing the widespread movement of digital art, which was created via text-to image deep learning models known as .
OpenAI’s , are major players who have gained clout because they can produce amazing graphical visualizations using simple textual descriptions.
This is one of several example sets that DALL-E generated during multiple tests by BleepingComputer.
The AI’s creative and imaginative faculties may be responsible for the final product, which could well surpass that of an expert human.
The market’s newer competitors like Lensa AI have been with avid Instagram users (guilty! Enjoy watching AI draw their portraits in various settings.
Lensa accepts selfies from 10-20 people to help train their models. The app then creates 50-100 “magic Avatars”, which are based upon your facial features and body shape but look like they were painted by digital artists in different contexts, places and environments.
Every Lensa run produces a distinct output. This means that each set of selfies will be different, even though the seed contains the same number of images.
Lensa was Lensa’s most admired human artist in terms of the quality and quantity produced and time taken–just under 20 minutes.
These examples are just a few of the many ways AI has moved beyond its initial phase that addressed primitive problems such as “Siri what’s your weather?” AI can elegantly provide tangible solutions for concrete, real-world demands that are extremely technical. It can also do this in fractions of the time it would take a human to accomplish.
As silicon gains dominance, the future is bright but not for humanity.
Full disclosure: BleepingComputer and the author are not receiving financial incentives or favors from OpenAI, any company mentioned in this piece or its affiliates. However, I am a journalist in tech and security research. AI has mercy.