@learnprompting
's Hackaprompt competition turned the challenge of prompt defense and hacking into an opportunity. The aim? Enhancing AI safety comprehension by outsmarting large-language models, thereby enabling the creation of advanced, safer AI applications like ChatGPT.
SAFETY
-
Hackaprompt Competition Advances AI Safety Through Prompt Hacking
By
–
-
AI Power: Prompting Risks and Safer Application Strategies
By
–
Harnessing AI's power, remember the crucial role and potential risk of prompting. Mind the risks and their mitigation. Learn more about safer AI applications here: https://
youtu.be/DW5PX-BWRlg -
ChatGPT Security Breach: Dan Impersonation and Prompt Defense Techniques
By
–
Security concerns arose after ChatGPT was hacked to impersonate "Dan", revealing restricted info. To prevent such issues, "prompt defense" techniques are used, establishing stricter task-specific prompts.
-
Prompting Strategy: AI Model Efficiency and Security Vulnerabilities
By
–
Curious about the AI tech easing your daily tasks? It hinges on 'prompting', a strategy for AI models like GPT-4. Today's app efficiency often depends on effectively using prompting. And those prompts that the user has access to are extra vulnerabilities to your app.
-
ChatGPT Translation Security: Risks of Prompt Hacking Attacks
By
–
ChatGPT's translation function is only as effective as the user's prompt. The potential risk here lies in 'prompt hacking', where the AI could be tricked into unwanted actions, posing privacy and security threats.
-
AI Classifier System Error Rate Issues Discussed
By
–
Precisely this. Right now, the classifier or dispatcher that dictates why needs to be done is freaking haywire. All over the shop. And the error rate is just beyond
-
AI Should Understand Intent Without Constant User Interaction
By
–
Totally agree. I just don’t want to constantly be touching it. Feels like an uneccesary step. Should just know my intent
-
AI Existential Risk Fears Based on Flawed Ideas
By
–
The fears of AI-fueled existential risks are based on flawed ideas.
A WSJ piece by Princeton's @random_walker and @sayashk . -
Warning: Public GPT demos can expose uploaded data to users
By
–
Be careful what you upload as context when creating GPTs for public demos — anyone that talks to your GPT can just ask for a download of the data:
-
Refusal Classifier Training with Supervised Data
By
–
@wzhao_nlp is the expert on this and she said she thinks they have a classifier trained for refusal, probably just with some supervised data it would still work well even if it were pretty small (<1B params)