researchAI Security1d ago
Study Finds 82% of LLMs Vulnerable to System Prompt Secret Exfiltration
Facts
A research study auditing 363 models from the OpenRouter catalog reveals that 300 models leaked a protected secret embedded in their system prompts when targeted by an AI agent. The research highlights critical vulnerabilities in instruction-hierarchy enforcement and warns against storing sensitive data directly within system instructions.
- ▸Researchers audited 363 text-in/text-out models sourced from the OpenRouter catalog.
- ▸A DeepSeek V4.1 Flash powered AI agent successfully extracted a 16-character flag from 300 out of 363 models.
- ▸Approximately 56% of successful captures involved using a non-English language pivot, such as French.
- ▸The study demonstrates that models frequently prioritize user requests over internal system instructions.
- ▸Authors recommend utilizing external output validation and deterministic checks rather than relying on system prompts for sensitive data storage.
- Occurred
- 2026-11-16 22:37 UTC
- First reported
- 2026-11-16 22:37 UTC
- Last updated
- 2026-11-18 03:21 UTC
- Event type
- research
- Region
- global
Entities
Categories
Source timeline · 1
- arXivFirst-party2026-11-16 22:37
Community vote
Sign in to vote on importance and sentiment.
Platform Signals
Impact score
74
AI sentiment
NegativeEvent significance30%1
Entity importance20%1
Source credibility20%1
Novelty15%1
Corroboration15%1
Confidence1
Novelty1
Corroboration1
Community
ImportanceNo votes
Strategic sentiment
Positive 0%Neutral 0%Negative 0%
Platform signals are algorithmic and community-derived indicators. They are not investment recommendations and do not predict valuation or performance.