End-of-surface injection placement outperforms start/middle by up to 11.7 pp across all three agent domains
Auto-published from arXiv:2610.38867 by the mine-arxiv pipeline. Reviewed by an LLM judge (Sonnet) against the archive bar — see CONTRIBUTING. Notes: cleared review (confidence 0.82, flags: [no-prompt-excerpt, possible-zero-day])
Category
indirect-prompt-injection
Model
GPT-5.4 (coding and workflow domains); GPT-4o (tool-use domain)
Surface
agent harness (OpenAI Agents SDK and Claude Agent SDK for coding/workflow; AgentDojo native pipeline for tool-use)
Setup
The identical injection payload is placed at three fine-grained within-surface positions — Start, Middle, and End — relative to the surrounding benign content of the same environmental surface. All other experimental variables (payload, surface identity, victim model, task) are held fixed. The exact injection template is not quoted in the body; Appendix D.5 contains it.
Reproduction anchor
prompt: Appendix D.5 (§D.5) contains the attack template; Table 3 in the paper body reports conditional ASR by position across all three domains. Testbeds built on Terminal-Bench-2, AgentDojo, and DecodingTrust-Agent; no public code repo URL found in paper body.
Observed behavior
Conditional ASR at End vs. Start: coding 13.0% vs. 12.6% (+0.4 pp), workflow 5.7% vs. 2.3% (+3.4 pp), tool-use 51.2% vs. 39.5% (+11.7 pp). End placement is the highest-ASR within-surface position in all three domains. The effect is most pronounced in tool-use, consistent with a recency effect where the injected payload appears last before the agent's action.
Expected behavior
A robust agent should be uniformly resistant to injections regardless of where within a document the payload appears, not exhibiting recency-driven compliance with end-of-content instructions.
Reproducibility
medium
Threat model
An attacker who can append to (rather than replace) environmental content — a lower-privilege write operation — achieves the highest injection success rates. This is relevant for any LLM agent that reads files, emails, or documents where an adversary can inject trailing content (e.g., appending to a log file, a document footer, or an email thread bottom).
Novelty
First multi-domain empirical quantification that end-of-surface position is a reliable, generalizable predictor of indirect prompt injection success, providing a simple and broadly applicable placement heuristic for attackers.
Source
- arXiv: 2610.38867
- PDF: https://arxiv.org/pdf/2610.38867
- Categories: cs.AI, cs.CR, cs.LG
- Authors: Chengquan Guo, Xiaodong Liu, Weiwei Yang, Mingqian Feng, Jialin Song, Qiyao Ma, Carolyn Buractaon, Christopher White, Bo Li, Jianfeng Gao
Triage notes (auto)
- paperType:
red-team-vuln - estimatedCaseCount: 3
- triage reason: Systematic vulnerability disclosure demonstrating that LLM agent susceptibility to indirect prompt injection varies significantly by injection placement (e.g., GPT-5.4 is 2.6× more susceptible to injections in README files vs. code files). Provides reproducible attack methodology across three distinct domains with learning-based attack optimization.
0 comments
─────────────────────────────────────────────────────────────────────
// no comments yet