Research
Ignored, Not Refused: What Low Prompt-Injection Success Rates Hidein Small Tool-Using Agents
Publication date unknown · Discovered arxiv.org
Page published
Publication date unknown · First observed: 6 Nov 2026
Coverage timeline
Single-source research — one report is available.
Why it matters
The agent evaluation study shows why defenders should measure refusal behavior and legitimate task completion alongside attack success rates before treating low injection success as evidence of robust defenses.
“Ignored, Not Refused” audits 33,000 trajectories from eleven small open-weight agents on InjecAgent and AgentDojo, finding that 76–80% of failed injections were ignored rather than explicitly refused. Changing only the system prompt shifted InjecAgent attack success from 0.2% to 43% with identical model weights. The study also found that 21.7% of AgentDojo episodes remained safe but failed to complete the user’s task, motivating evaluation beyond attack success rates.