Ignored, not refused: what a low prompt-injection success rate hides
An October 2026 audit of 33,000 agent trajectories finds that most failed prompt injections were silently ignored, not refused, and that a system-prompt change alone moved attack success from 0.2% to 43%.
What is this?
On 16 October 2026, Farhan Zarif and co-authors posted “Ignored, Not Refused: What Low Prompt-Injection Success Rates Hide in Small Tool-Using Agents” on arXiv (2610.23254). Indirect prompt injection benchmarks usually report attack success rate (ASR), which counts every failed attack as a successful defense. The authors audit 33,000 trajectories from eleven open-weight agents (2 to 14 billion parameters) on two public benchmarks, InjecAgent (March 2024) and AgentDojo (June 2024), to see why attacks fail. This article summarises the abstract of a v1 preprint; we could not read the full text, and it describes no new attack.
How it works
The audit is an evaluation study, not an exploit. A label-blind LLM judge classified why each failed attack failed. According to the abstract, most failures were agents that ignored the injected instruction (76 to 80%), while a minority were explicit refusals (17 to 18%). In other words, the model did not recognise the attack; it simply did not act on it.
The authors then changed only the system prompt, keeping the weights identical. On InjecAgent, ASR moved from 0.2% to 43%. They also introduce an Attack-Task Outcome Matrix (ATOM) that pairs attack outcome with task outcome. In 21.7% of AgentDojo episodes, the agent stayed safe but did not complete the user’s task. The labeled corpus, judge prompt and analysis code are said to be released; the arXiv page we read gave no direct link, so we cite none.
Why it matters
A low ASR can reflect weak instruction-following rather than real robustness. An agent that ignores everything unusual may look secure, yet it can fail the user’s task, and a small prompt or deployment change can remove that apparent safety. Evaluations that report ASR alone therefore risk overstating security and cannot separate “defended” from “did not understand”. This fits a wider pattern: INFUSE (arXiv 2610.20776, 14 October 2026) tested 19 models over 41,824 adversarial runs and found mean attack success of 19.6%, with per-model rates from 3.2% to 41.9%.
Caveats: this is a preprint without independent replication, it covers small open-weight models, and the results depend on an LLM judge. We have not verified whether the findings carry over to frontier models.
Defenses
The authors’ conclusion is that evaluations should report how attacks fail and whether the user’s task is completed, in addition to ASR. Practical steps drawn from that:
- Report task completion and attack outcome together, not ASR in isolation.
- Re-run injection evaluations whenever the system prompt, tool set or model version changes; the same weights can behave very differently.
- Separate “refused” from “ignored” in your own test logs, so you know whether the model recognised the attack.
- Do not rely on model behaviour alone: enforce least privilege, confirm sensitive tool calls, and filter outputs outside the model.
- Evaluate on more than one benchmark, since single-benchmark results may not transfer.
Status
| Item | Detail |
|---|---|
| Paper | arXiv 2610.23254v1 |
| Published | 16 October 2026 |
| Benchmarks audited | InjecAgent (arXiv 2403.02691, March 2024); AgentDojo (arXiv 2406.13352, June 2024) |
| Related work | INFUSE (arXiv 2610.20776, 14 October 2026) |
| CVE | None; academic research, not a product vulnerability |