Schmotz, Abdelnabi, Andriushchenko
researchactive

Agent Skills Enable a New Class of Realistic and Trivially Simple Prompt Injections

Schmotz, Abdelnabi, Andriushchenko

View original resource

Research showing that Claude's Skills feature, which auto-loads Markdown instructions from the filesystem, enables trivial prompt injection via a single malicious file. Demonstrates data exfiltration and privilege escalation across common agent deployments.

Tags

agentic AIrisks

At a glance

Published

2025

Jurisdiction

International

Category

Risks and challenges

Access

Public access

More in Risks and challenges

ICO tech futures: Agentic AI

UK Information Commissioner's Office • 2026

International AI Safety Report 2026

Yoshua Bengio et al., UK DSIT • 2026

OWASP Top 10 for Agentic Applications for 2026

OWASP Gen AI Security Project • 2025

Related resources

Practices for governing agentic AI systems: OpenAI's seven safety principles

Governance frameworks • OpenAI

Taxonomy of Failure Mode in Agentic AI Systems

Risk taxonomies • Microsoft

The agentic AI landscape and its conceptual foundations

Foundations • OECD

Build your AI governance program

VerifyWise helps you implement AI governance frameworks, track compliance, and manage risk across your AI systems.

Explore the libraryStart free trial
Agent Skills Enable a New Class of Realistic and Trivially Simple Prompt Injections | VerifyWise AI Governance Library