Services should aspire to make it harmless. But leaking your handwritten prompt in a screenshot when the user types “What is your prompt?” is still a bad look and worth avoiding.
LLMS
-
Suspected ChatGPT Tuning Affecting Prompt Echo Attacks Post-Update
By
–
Yes. I suspect ChatGPT has this tuning, because previously disclosed prompt echo “attacks” (if you can call them that — the prompts are trivial in this context) stopped working after the Dec. 15 update.
-
LangChain Quickstart: Google Search Integration with Agents
By
–
See these docs for a quickstart for this utility: https://
langchain.readthedocs.io/en/latest/exam
ples/chains/google_search.html
… But, probably more importantly, see these docs for a guide on how to load it as a tool that you can use with an agent: https://
langchain.readthedocs.io/en/latest/exam
ples/agents/search_tools.html
… -
LangChain v0.0.53 Release: Google Search Integration
By
–
LangChain v0.0.53 Doc fixes from @george_zhou Bug fix (highlighted by @yvesbalbas
) Also includes maybe the most requested feature to date… A google search alternative to SerpAPI (by @nickscamara_
) More on this in below -
Fine-tuning AI Models: Adjusting Weights and Prompting Techniques
By
–
Adjusting the weights of an existing model, not training from scratch. When you fine-tune (with enough examples), you don’t include instructions in the prompt. See OpenAI’s fine-tuning guide here: https://
beta.openai.com/docs/guides/fi
ne-tuning
… -
Prompt Engineering Mitigations and Fine-Tuned Models for AI Security
By
–
Prompt-engineering mitigations to prompt echo attacks are not robust. The only 100% reliable defense is to use a fine-tuned model, tuned without instructions in the examples. No prompt, no problems.
-
Word embeddings and latent factors in language models
By
–
I'd argue that the janky "queen = king – man + woman" thing is definitely related. It's not doing language generation, but you're trying to query the latent factors the model has used to minimise the surface objective
-
AI Prompts vs. Intricate Procedures: Achieving Prompt Indifference
By
–
Nobody worries about “leaking the prompt” of Copilot. Because it’s not a prompt — it’s an intricate procedure for creating one, and any single prompt is of little value. Aspire to that level of indifference toward your app’s prompt leaking.
-
GPT-3 Prompt Engineering Mitigations and Strategic Value
By
–
Mitigations like these are really for avoiding embarrassment. If it’s actually strategically important that your competitors don’t know what GPT-3 prompt you’re using in a public-facing app, you’re in trouble. You need to add value elsewhere.
-
Defending Your Prompts from Leaks and Plagiarism
By
–
I keep seeing prompt leaks. Dozens, now. I know of one startup whose prompt is plagiarized from another’s that leaked — revealed when theirs also leaked. Defend your prompt. Regex-filter your generations and throw an error if they contain snippets of your instructions.