GCD AI RADAR
incidents/AIR-2026-1401

Ignored, Not Refused: Analyzing Indirect Prompt Injection in Agents

Attack on AI1 Research / proof of conceptPrompt injectionAgent hijack

An audit of agent execution trajectories reveals that most failed indirect prompt injections are ignored rather than actively refused by small tool-using models. The findings highlight subtle safety vulnerabilities in agentic tool use.

Details

type
Attack on AI
An AI model, agent or AI-powered product was the thing exploited, abused or leaked.
categories
  • Prompt injectionprimary
    Untrusted content (web pages, documents, emails, tool output) steers a model into ignoring its instructions.
  • Agent hijack
    An AI agent is tricked into unauthorized actions: tool misuse, data exfiltration through the agent, or acting against its principal.
owasp llm top 10
mitre atlas
Both rows are derived from the type and categories (OWASP 2025 edition, ATLAS release 2026.09), not classified per record. How
severity
1 · Research / proof of concept
Demonstrated by researchers; no known real-world exploitation.
targets
Open-weight model, Framework / tooling
products
InjecAgent, AgentDojo
incident date
19 Oct 2026
first reported
19 Oct 2026
origin
News & research feeds
classifier confidence
0.90 · published at or above 0.80; see the methodology
tags
prompt-injection · indirect-prompt-injection · agents
record
created 2026-10-19 05:07Z · updated 2026-10-19 05:07Z · json

Sources · 1

  1. 2026-10-19
    Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents
    arXiv · arxiv.org/abs/2610.23254

Titles and links belong to their publishers. The summary above is written by GCD AI RADAR in its own words; data is CC BY-SA 4.0.

Related incidents