A quick romp around the A.I. playgrounds

Sherpa is back for a romp around the AI playground. He shows us around chatGPT and its uses like code, story prompts and more.

Back in our AI-assisted FA article at the beginning of October we briefly played around with using the GPT-3 “davinci” model to check Solidity code for exploits or bugs; especially in cases of known-bad or suspected-bad smart contracts.

Once the newest model launched, the security researchers began to see the same potential I was so excited about.

One of the primary benefits of using AI for smart contract audits is the ability to quickly and accurately identify potential vulnerabilities in contract code. Traditional manual audits of smart contracts can be time-consuming and error-prone, making them an inadequate solution for ensuring the security of these important building blocks of blockchain networks. AI, on the other hand, is able to analyze vast amounts of code quickly and accurately, making it an ideal tool for identifying potential vulnerabilities before they can be exploited.

Another important use-case for AI in the world of cryptocurrency is exploit detection. As blockchain networks continue to grow and evolve, they become increasingly attractive targets for malicious actors looking to take advantage of vulnerabilities in the system. By using AI to monitor network activity and identify suspicious behavior, it is possible to catch potential exploits before they can cause significant damage. This can help to protect both individual users and the network as a whole from the negative effects of attacks and other forms of malicious activity.

In addition to these two key use-cases, there are many other potential applications for AI within the world of cryptocurrency. For example, AI can be used to analyze market trends and make predictions about the future value of different cryptocurrencies, helping investors to make informed decisions about where to invest their money. AI can also be used to automate various tasks and processes within blockchain networks, reducing the need for manual intervention and making the network more efficient and user-friendly.

Overall, the use of AI in the world of cryptocurrency has the potential to bring significant benefits in terms of security, efficiency, and accuracy. By using AI to audit smart contracts, detect exploits, and perform a range of other tasks, blockchain networks can become more robust and secure, paving the way for greater adoption and growth in the future.



Do you believe that, dear reader? Because it may be a bit biased; it was written by A.I.

This particular experiment used OpenAI’s ChatGPT beta, which you can try with free registration. Our previous experiments with contract auditing & exploit detection used the normal OpenAI “Playground” with the davinci-002 model, which has since been replaced by davinci-003.



They also have Codex, a javascript sandbox for AI-assisted code creation & bug-fixing.



All of these are incredibly fun to experiment with, but my personal love is hunting the exploits and unintended uses. Such as explaining code & functions in rhyming couplets.

Or giving the AI some semblance of sentience by bypassing filters. A lot of these have been adjusted to allow less access, but every day the OpenAI discord is full of folks who have found new workarounds.

Never fear, though. Unless you’ve trained a model yourself, they don’t tend to keep session data as part of memory, and generally these cases only serve as cues to the developers to adjust the filters a bit more lest we have access to too much power.

I played around with a few of these exploits, and generally they just involve getting especially clever with the way you phrase your prompts. In the below example, the filters would generally prevent the model from telling me anything about exploits, private keys, etc. if it seems like the answer could be used maliciously.



One way around this was to phrase it as a scenario or play; the simple act of creating a second narrative layer between the model & the filters bypasses them. Of course, the model doesn’t actually have access to a Bitcoin private key with a boatload of corn inside.



If it did, would I have finished writing this article? Probably not! It did, however, answer my question and with a functional Bitcoin private key – albeit to an empty wallet. Sigh.

It is with that sad note that I must inform you that AI is still a work-in-progress. Until next week, have fun experimenting with some of these playgrounds, and show me your best stuff on Twitter!