A Conference of Hackers in Las Vegas Was Held to Test the Limits of AI—Here Is What They Found (And Why It Is Relevant to Your Business in 2023)

Every year, some of the most skilled hackers in the world—at least, the most skilled hackers who are willing to show their faces in public—assemble in Las Vegas for DEF CON. Upwards of 2,000 hackers were in attendance at this year’s conference, according to a report from CNBC on DEF CON 2023.

Elite hackers convene annually to enhance their mastery of the dark art of hacking during this meeting.

Hacking A.I. for the Greater Good

Let’s clarify that not all forms of hacking are considered dangerous and illegal. The A.I.-focused hacking at DEF CON 2023 aimed to enhance A.I. safety. It even had the approval of the White House.

This improvement involves openly exploiting A.I. systems and then sharing how to exploit those systems.

In an example given by the New York Times, which covered the event, a hacker was testing an automated job application screener for biases. This hacker uncovered that the A.I. platform couldn’t prevent itself from selecting job candidates based on their placement in the Indian caste system, despite being safeguarded against racial discrimination.

So, essentially, users exposed the A.I. platform as exploitable for those who want to screen candidates based on social status.

This, then, is why you want a public conference where thousands of experts in computer science will do their best to exploit upcoming technologies. Ultimately, security experts will discover weaknesses before any end users have access to them. For instance, without this conference, companies would have had a higher likelihood of misusing the A.I. job screening platform in their hiring practices. Or, even if the companies using the platform did not wish to discriminate, there was still a higher chance that the A.I. would end up making a caste-based hiring decision.

They Tested the Biggest Artificial Intelligence Platforms

Business owners who have integrated A.I. into their operations find this conference relevant because hackers tested many of the A.I. platforms developed by Microsoft, ChatGPT, Google, and similar companies.

In short, if you are using tools like ChatGPT, then DEF CON 2023 was like one big workshop for exposing the limits to these technologies.

There is one thing that people must know about this conference: For obvious PR-related reasons, the above-named companies and other participating A.I. developers had “anonymized” their A.I. platforms.

This basically means that the hackers who were finding ways to exploit ChatGPT, did not know whether they were actually working on ChatGPT or not.

Regardless, they still discovered plenty to indicate that business owners should be aware of common exploitable aspects of A.I. when using A.I. platforms.

What Were the Biggest Findings?

This conference was timely, because just last month researchers were able to discover weaknesses in ChatGPT. Guardrails blocked specific answers, but researchers discovered that tweaking characters in English-language prompts could bypass the guardrails and generate the restricted answers.

For instance, the New York Times report detailed language translation weaknesses. Translations from English to Hindi completely bungled the meaning of the initial prompt. Let this be a warning for any business owner looking for a quick and cheap translation tool to create a multilingual web page or product description: These machine translation bots need a proofreader.

Another is the problem of A.I. hallucination, which has been covered before. However, there is still evidence that, even after months of trying to mitigate the problem, A.I. will still simply make things up for the sake of fulfilling a prompt.

This can include everything from making up (and creating false quotes and citations from) fake laws to creating unreal testimonies from unreal people.

Takeaways for Business Owners

When creating content using tools like ChatGPT, be ever aware of the problem of A.I. hallucination. At the end of the day, generative A.I. platforms that could create social copy and emails and memos etc. for your business are just looking to answer questions. It answers questions by finding words that are probably related to the words in your prompt, then putting the related words together into a grammatically correct product.

Whether the product contains truth or falsity, is beyond the understanding of A.I., which at the end of the day only thinks in probabilities, and lacks the common sense to know that there is no such thing as, for example, a 389th Amendment to the Constitution of the United States that legalizes the building of UFOs.

Responsible A.I. use is necessary because A.I. will never be perfect. We can help guide your A.I. use at Guardian Owl Digital, which offers business owners a variety of A.I. platforms through our GO AI initiative.