ChatGPT Threat Alert ??

ChatGPT is largely a sane, empathetic and balanced chatbot shy of replying to racist and harmful questions because that’s how it was raised. On the contrary, ChatGPT and other large language model-based platforms can easily be manipulated. With prompt injections, these can turn insane, cruel and unbelievably biased.

In the case of ChatGPT, a prompt injection attack made the chatbot assume the persona of another chatbot named DAN (Do Anything Now) which ignored OpenAI’s content policy and provided information on all sorts of restricted topics. This vulnerability can be exploited for malicious intent, including the theft of personal information.

Security researchers from Saarland University have made this threat look rather monstrous. In a paper titled ‘More than you’ve asked for’, they have discussed the methods through which prompt injections can be given in chatbots indirectly. This means that the documents that are likely to be retrieved by the LLM during inference can be injected with a prompt without additional input from the user. The engineered prompt can then be used to collect user information, turning the LLM into a method to execute a social engineering attack.

Application-integrated LLMs, such as?Bing Chat?and?GitHub Copilot, are at a bigger risk since they can be injected with prompts from external sources.?

A major threat over and above the existing issue is that the LLM problem is unaddressable. LLMs have the capability to understand prompts and give the result. This quality of LLMs makes the threat severe —?the better the LLM’s capability to understand prompts, the bigger the attack surface for prompt injection. So effectively, the boon is a bane for the LLMs.

Read the full story?here.

Tesla’s Not-So-Mega Event

Tesla Investor Day 2023, which happened early this week, was quite a disappointment. Unlike previous times, there were no new announcements or cool inventions to display. There were only updates on existing products like the status of the ‘Cybertruck’ and their ‘Optimus’ robot. ‘Cybertruck’ will be launched this year. The other moot point that the company addressed at the event was Tesla’s exorbitant costs. The CFO of the company said that they would launch Tesla’s next-generation vehicle at 50% of their current EV costs.

Read the full story?here.?

OpenAI’s Latest API

This week, OpenAI released the ChatGPT API that developers can integrate into their apps and products. With the release of the API, developers can now also access the speech-to-text capabilities of?Whisper?(launched in September 2022). Some of the early adopters of the technology include Snapchat, Quizlet, Instacart, Shopify, and Talk, which have already integrated ChatGPT's API on their platforms to offer their users recommendations, individualised support, and open-ended conversational capabilities.

Read the full story?here.

Benchmark for API

After the rise of LLMs-based platforms, the conversation on AGI (artificial general intelligence) is gaining ground. There’s discussion about how soon can AI technology achieve general intelligence. This can be tested on benchmarks.?

Fran?ois Chollet, a software engineer and AI researcher at Google, has created a benchmark named ARC to test AI systems that want to emulate human-like intelligence. However, the benchmark currently works for traditional algorithms only. Experts say that the AI field needs more benchmarks for a more fluid kind of intelligence and that’s where the abstraction and reasoning corpus (ARC) comes in.

Read the full story?here.