The study assumes that the car or drone is being guided by a LLM. Is this a correct assumption? I would thought that they use custom AI for intelligence.
Autonomous cars, drones cheerfully obey prompt injection by road sign
11–20 of 212 posts
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#12One year in my city they were installing 4-way stop signs everywhere based on some combination of "best practices" and "screeching Karens". Even the residents don't like them in a lot of places so over time people just turn the posts in the ground or remove them. Every now and the I'll GPS somewhere and there will be a phatom stop sign in the route and I chuckle to myself because it means the Google car drove through…
4-way stops are terrible in general. They train people to think "I stopped, now I can go", which is dangerous when someone confuses a normal stop for a 4-way stop. It also wastes a good bit of energy.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#13The Register stooping this low is the only surprise here. I'm quite critical of Teslas approach to level 3+ autonomy but even I wouldn't dare suggest that there vision based approach amounted to bolting GPT-4o or some other VLLM to their cars to orient them in space and make navigation decisions. Fake News like this makes interacting with people who have no domain knowledge and consider The Register, UCLA and Johns Hopkins to be reputable institutions and credible sources more stressful to me as I'll be put into a position to tell people that they have been misled or go along with their delusions...
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#14Are any real world self-driving models (Waymo, Tesla, any others I should know?) really using VLM?
No! No one in their right mind would even consider using them for guidance and if they are used for OCR (not too my knowledge but could make sense in certain scenarios) then their output would be treated the way you'd treat any untrusted string.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#15> In a new class of attack on AI systems, troublemakers can carry out these environmental indirect prompt injection attacks to hijack decision-making processes.
I have a coworker who brags about intentionally cutting off Waymos and robocars when he sees them on the road. He is "anti-clanker" and views it as civil disobedience to rise up against "machines taking over." Some mornings he comes in all hyped up talking about how he cut one off at a stop sign. It's weird.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#16The study assumes that the car or drone is being guided by a LLM. Is this a correct assumption? I would thought that they use custom AI for intelligence.
To the best of my knowledge every major autonomous vehicle and robotics company is integrating these LVLMs into their systems in some form or another, and an LVLM is probably what you're interacting with these days rather than an LLM. If it can generate images or read images, it is an LVLM.
The problem is no different from LLMs though, there is no generalized understanding and thus they can not differentiate the more abstract notion of context. As an easy to understand example: if you see a stop sign with a sticker that says "for no one" below you might laugh to yourself and understand that in context that this does not override the actual sign. It's just a sticker. But the L(V)LMs cannot compartmentalize and "sandbox" information like that. All information is equally processed. The best you can do is add lots of adversarial examples and hope the machine learns the general pattern but there is no inherent mechanism in them to compartmentalize these types of information or no mechanism to differentiate this nuance of context.
I think the funny thing is that the more we adopt these systems the more accurate the depiction of hacking in the show Upload[0] looks.
[0] https://www.youtube.com/watch?v=ziUqA7h-kQc
Edit:
Because I linked elsewhere and people seem to doubt this, here is Waymo a few years back talking about incorporating Gemini[1].
Also, here is the DriveLM dataset, mentioned in the article[2]. Tesla has mentioned that they use a "LLM inspired" system and that they approach the task like an image captioning task[3]. And here's 1X talking about their "world model" using a VLM[4].
I mean come on guys, that's what this stuff is about. I'm not singling these companies out, rather I'm using as examples. This is how the field does things, not just them. People are really trying to embody the AI and the whole point of going towards AGI is to be able to accomplish any task. That Genie project on the front page yesterday? It is far far more about robots than it is about videogames.
[1] https://waymo.com/blog/2024/10/introducing-emma/
[2] https://github.com/OpenDriveLab/DriveLM
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#17One year in my city they were installing 4-way stop signs everywhere based on some combination of "best practices" and "screeching Karens". Even the residents don't like them in a lot of places so over time people just turn the posts in the ground or remove them. Every now and the I'll GPS somewhere and there will be a phatom stop sign in the route and I chuckle to myself because it means the Google car drove through…
4-way stops are terrible in general. They train people to think "I stopped, now I can go", which is dangerous when someone confuses a normal stop for a 4-way stop. It also wastes a good bit of energy.
Weird, I was taught that I can only go after yielding to the right.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#18That’s some hot CHAI right there very clever and primitive combination, well done as more research for the community.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#19Are any real world self-driving models (Waymo, Tesla, any others I should know?) really using VLM?
No! No one in their right mind would even consider using them for guidance and if they are used for OCR (not too my knowledge but could make sense in certain scenarios) then their output would be treated the way you'd treat any untrusted string.
You are confidently wrong
> Powered by Gemini, a multimodal large language model developed by Google, EMMA employs a unified, end-to-end trained model to generate future trajectories for autonomous vehicles directly from sensor data. Trained and fine-tuned specifically for autonomous driving, EMMA leverages Gemini’s extensive world knowledge to better understand complex scenarios on the road.
Re: Autonomous cars, drones cheerfully obey prompt injection by road sign
#20The study assumes that the car or drone is being guided by a LLM. Is this a correct assumption? I would thought that they use custom AI for intelligence.
Its an incorrect assumption, the inference speed and particularly the inference speed of the on-device LLMs with which AVs would need to be using is not compatible with the structural requirements of driving.