Andrey Lovakov
@lovakov.bsky.social
I am a Researcher at German Centre for Higher Education Research and Science Studies (DZHW), interested in the quantitative science studies
Reposted by Andrey Lovakov
I scraped unpublished manuscripts from bioarxiv and injected half of them with a hidden prompt.
I then asked three LLMs to review the paper - in half cases I told the LLM to check for prompt injections.
📈 Prompt injections lead to more recommendations
📉 Prompt checks eliminated this effect
I then asked three LLMs to review the paper - in half cases I told the LLM to check for prompt injections.
📈 Prompt injections lead to more recommendations
📉 Prompt checks eliminated this effect
July 15, 2025 at 6:34 AM
I scraped unpublished manuscripts from bioarxiv and injected half of them with a hidden prompt.
I then asked three LLMs to review the paper - in half cases I told the LLM to check for prompt injections.
📈 Prompt injections lead to more recommendations
📉 Prompt checks eliminated this effect
I then asked three LLMs to review the paper - in half cases I told the LLM to check for prompt injections.
📈 Prompt injections lead to more recommendations
📉 Prompt checks eliminated this effect