Security
ICML-listed research finds prompt-injection defense tradeoffs
Researchers report that no tested model-defense configuration achieved both high security and high fidelity in a benchmark of indirect prompt-injection attacks.
International Conference on Machine Learning
What happened
ICML’s 2026 downloads index lists a paper titled “Security–Fidelity Tradeoffs: No Universal Defense Against Prompt Injection.” The conference index does not give an individual presentation or publication date.
In an accessible preprint dated June 29, the researchers evaluated 1,168 examples across 48 model-defense configurations. They reported that none achieved both high security and high fidelity, which measures whether a defense preserves intended behavior on non-malicious tasks.
The reported endpoints were 96.5% fidelity with 47.8% security, and 99.3% security with 71.0% to 73.9% fidelity. Prompt injection is an attack in which untrusted input is combined with higher-trust application instructions.
Why it matters
Systems that connect language models to documents, tools, or outside content may need to balance blocking malicious instructions against preserving useful model behavior. The findings provide measured evidence of that tradeoff in the researchers’ tested setting.
What remains unclear
- The evaluation covers the paper’s indirect-prompt-injection tasks, models, defenses, metrics, and operating conditions. It does not show that every language model or attack is insecure, or that future defenses cannot improve the tradeoff.
- The official ICML index confirms the paper listing, but its linked virtual-poster page was not accessible and the index provides no individual-paper presentation date.
Sources
-
Downloads
Used for: Confirmation that ICML’s 2026 index lists the paper.
-
Security--Fidelity Tradeoffs: The Hidden Cost of Prompt Injection Defense
Used for: The benchmark scope, configuration count, and reported security-fidelity results.
-
prompt injection - Glossary
Used for: Definition of prompt injection.