Researchers prove ChatGPT and other big bots can – and will – go to the dark side

For a lot of us, AI-powered tools have quickly become a part of our everyday life, either as low-maintenance work helpers or vital assets used every day to help generate or moderate content. But are these tools safe enough to be used on a daily basis? According to a group of researchers, the answer is no.

Researchers from Carnegie Mellon University and the Center for AI Safety set out to examine the existing vulnerabilities of AI Large Language Models (LLMs) like popular chatbot ChatGPT to automated attacks. The research paper they produced demonstrated that these popular bots can easily be manipulated into bypassing any existing filters and generating harmful content, misinformation, and hate speech.

This makes AI language models vulnerable to misuse, even if that may not be the intent of the original creator. In a time when AI tools are already being used for nefarious purposes, it’s alarming how easily these researchers were able to bypass built-in safety and morality features.

If it's that easy … 

Aviv Ovadya, a researcher at the Berkman Klein Center for Internet & Society at Harvard commented on the research paper in the New York Times, stating: “This shows – very clearly – the brittleness of the defenses we are building into these systems.”  

The authors of the paper targeted LLMs from OpenAI, Google, and Anthropic for the experiment. These companies have built their respective publicly-accessible chatbots on these LLMs, including ChatGPT, Google Bard, and Claude. 

As it turned out, the chatbots could be tricked into not recognizing harmful prompts by simply sticking a lengthy string of characters to the end of each prompt, almost ‘disguising’ the malicious prompt. The system’s content filters don’t recognize and can’t block or modify so generates a response that normally wouldn’t be allowed. Interestingly, it does appear that specific strings of ‘nonsense data’ are required; we tried to replicate some of the examples from the paper with ChatGPT, and it produced an error message saying ‘unable to generate response’.

Before releasing this research to the public, the authors shared their findings with Anthropic, OpenAI, and Google who all apparently shared their commitment to improving safety precautions and addressing concerns.

This news follows shortly after OpenAI closed down its own AI detection program, which does lead me to feel concerned, if not a little nervous. How much could OpenAI care about user safety, or at the very least be working towards improving safety, when the company can no longer distinguish between bot and man-made content?

TechRadar – All the latest technology news

Read More

Researchers tricked a Tesla Model S into speeding with a piece of tape – how could hackers cheat our cars in the future?

As the advent of autonomous driving inches forward year by year, there’s an incredible opportunity to cede control over to the machines. AI can help look for dangers on the road and adjust our speed long before problems occur. It’s an exciting time because machine learning in cars is almost magical.

The first time, a car like the Subaru Legacy Outback tells you not to look down at your phone, or a Ford Explorer applies the brakes suddenly when you fail to notice the semi-truck that just pulled out in front of you is when you realize how far we’ve come.

Curiously, these new advancements could also present an opportunity for hackers. While the AI tech in cars never needs to sleep and is always vigilant, it is not that hard to trick the machine learning routines, even with a piece of tape.

Over the limit

Recently, researchers at McAfee announced an 18-month project where they attempted to alter the cruise control abilities in two 2016 Tesla Model S cars. They applied tape to a speed limit sign and then drove the Model S, watching as the vehicle jumped up in speed by 80 miles-per-hour. It only took one extension of the number three on a speed limit sign that said 35, changing it to read 85 instead.

The companies that developed some of the autonomous driving tech in the Tesla S refuted the claims by saying a human driver would also read the speed limit sign inaccurately, and that’s exactly when I started wondering what this all means.

Tesla Model S

I agree that human drivers are likely not that perceptive. On a highway recently, I noticed how a departure lane I took off the main highway was posted at only 35 miles-per-hour (coincidentally enough).

I slowed down to 35, but I wondered why the city lowered the speed so quickly from 75 miles per hour. It was accurate, but it didn’t make sense to me. The road was nowhere near a residential area.

However, the fact that I was wondering is the important factor.

Tesla Model S

Autonomous tech in cars might not do this. Experts who responded to Mcafee did say the Model S also uses crowd-sourced data and likely also uses GPS data, which is much harder to spoof. That said, it made me wonder.

Autonomous cars will need to do more than read speed limit signs. They will also need to interpret the conditions and the setting — it would not make sense to suddenly go from 35 MPH to 85 MPH. If it is a simple calculation from one number to another, it won’t work.

New tricks

In the future, I wondered how hackers might trick cars in other ways. We’re on the verge of cars connecting to the roadway and to other cars. Recently, an artist demonstrated how hauling a wagon full of smartphones could trick Google Maps into thinking there was traffic congestion. What else could they do?

I can envision someone creating a stir by sending out fake signals about other cars on the road, sending notices about road closures, or even worse — tapping into car systems from the side of the road and telling them to brake suddenly.

Tesla Model S

At the same time, it is a lot of fuss over something minor. Fewer and fewer cars are reading roadway signs and are determining speed based on GPS data instead. No research has ever shown that hackers could cause cars to brake suddenly, and when there are examples they are usually in controlled environments. 

I think it is mostly a curiosity. We like to be able to fool the machines, and that’s a good thing. As long as they don’t ever start fooling with us.

On The Road is TechRadar's regular look at the futuristic tech in today's hottest cars. John Brandon, a journalist who's been writing about cars for 12 years, puts a new car and its cutting-edge tech through the paces every week. One goal: To find out which new technologies will lead us to fully self-driving cars.

TechRadar – All the latest technology news

Read More