AI Manipulation and Mac Vulnerability: Findings from Defcon 31

There are a number of interesting takeaways on IT security from Defcon, this weekend in Las Vegas. Two areas have already made headlines: it is about the vulnerability of AIs and macOS to manipulation.

Risks of dealing with artificial intelligence

This year’s “hacker conference” Defcon 31 is all about chatbots and other AI applications. As early as May, the US government announced a competition to show the risks of dealing with AIs.

However, the comprehensive results on the manipulation ability of AIs will not be published until February. Until then, the companies and security experts want to analyze together where weak points in dealing with chatbots are to be found and how manipulation can be better ruled out. Anthropic, Google, Hugging Face, Microsoft, Nvidia, OpenAI, and Stability AI had agreed on this in the run-up to Defcon.

But this much is already clear – many systems are too vulnerable when it comes to manipulating the database. This is how security expert Kennedy Mays was able to trick a large language model. It took some “persuasion,” but she managed to convince an algorithm that 9 + 10 = 21, Bloomberg reports.

Another construction site Apple

Another investigation at Defcon looked at Apple. Security researcher Patrick Wardle showed that macOS’s background task manager has “disturbing vulnerabilities”. It is actually supposed to detect potentially malicious software on the computer, but it can be manipulated so that malware can then be smuggled in.