uipath-gov-access-policy
Pass
Audited by Gen Agent Trust Hub on May 5, 2026
Risk Level: SAFEPROMPT_INJECTION
Full Analysis
- [PROMPT_INJECTION]: The paper investigates 'prompt injection attacks' where malicious instructions are inserted into a model's context to divert it from its intended task. The study specifically analyzes the relationship between input length and the success rate of these attacks. The researchers tested models like Llama-3-8B and Phi-3-Mini-4k with varying context lengths (e.g., repeating task instructions multiple times). The results showed that increasing the number of instruction repetitions or adding unrelated 'filler' text generally reduced the success rate of the attacks. For instance, increasing instruction repetitions from 1 to 50 in Llama-3-8B-Instruct dropped the attack success rate from 27% to 4%. The study suggests that long-context LLMs become better at following instructions and filtering out noise as the input size increases, thus making them more resilient to simple injection attempts.
Audit Metadata