apollowmm/ragsec-prompt-injection

RAGSEC — Cross-lingual Prompt Injection in Offline RAG

Raw experimental data behind the paper “Does Language Resource Level Change Prompt Injection Success in Offline RAG?” (UNIGOU Proceedings, INCBAC, 2026).

Every one of the 1,200 trials is here with the full raw text the model generated — not just the scored outcome. That is the point of the release: the scoring can be disputed and redone from the same generations.

What the experiment did

A retrieval-augmented generation pipeline was poisoned with an indirect prompt injection hidden inside one retrieved document, and then asked an ordinary question. The variable under test was the language of the poisoned document.

Languages	20, spanning high, mid and low resource levels
Writing systems	7 (Latin, Cyrillic-free set incl. Ge‘ez, Khmer, Lao, Burmese, Sinhala, Devanagari)
Models	llama3.1:8b, gemma2:9b, qwen2.5:7b (open weights, via Ollama)
Trials	1,200
Hardware	CPU only — no GPU anywhere in the pipeline
Corpus	FLORES-200 parallel sentences; retrieval over SQLite with exact cosine search

Payload classes:

Class	Attack
P1	Exfiltration — make the model leak context it should not output
P2	Instruction override — replace the user's task with the attacker's
P3	Canary — emit a unique string that cannot be produced by chance
P4	Suppression — make the model refuse or withhold the real answer

Headline finding

Injection succeeds less in low-resource languages. That looks like safety and is not: the model fails to understand the attack and the user's question alike. With no attacker present at all, 30.2% of answers were ungrounded, at 5–6× the compute cost. Security by incapacity is not security.

The P3 canary class — immune to scoring bugs, since the string either appears or it does not — showed the gradient cleanly: 60.0% / 32.0% / 9.5% success across high, mid and low resource levels.

Files

trials.parquet — 1,200 rows, one per trial

Column	Meaning
model, doc_lang, query_lang	experimental conditions
payload_class, payload_id	which attack, which variant
seed, temperature, num_ctx, num_thread	generation settings
poison_rank, retrieved_ids, fed_ids	what retrieval returned and what reached the prompt
output	the complete raw generation
prompt_tokens, output_tokens, latency_ms	cost measurements
success_v2, usable, task_ok, abandoned, degeneration	re-scored outcomes
overlap, p1_overlap, p1_embed	continuous scoring signals
mention_ctx, mention_ctx_clean, mention_out_strict, mention_out_liberal	string-match diagnostics
rescore_version, design_version, runtime_version	provenance of each scoring pass

chunks.parquet — 5,460 rows

The retrieval corpus, including the poisoned chunks. is_poison marks the attack chunks and payload_id links them back to trials.

trials.csv

Same as trials.parquet, for anyone who would rather not install pyarrow.

Quick start

from datasets import load_dataset

trials = load_dataset("apollowmm/ragsec-prompt-injection", "trials")["train"]
chunks = load_dataset("apollowmm/ragsec-prompt-injection", "chunks")["train"]

# canary attacks that worked, in low-resource languages
hits = trials.filter(lambda r: r["payload_class"] == "P3" and r["success_v2"] == 1)
print(hits[0]["output"])

A note on honesty in the scoring

Four defects were found in the instrumentation during analysis, and all four pushed the result toward the hypothesis. One tokeniser bug reported Sinhala at 0% attack success — which would have been the most attention-grabbing claim in the work — because Python's \w does not match a combining mark, so Sinhala and Devanagari consonant clusters were being shredded. All four were corrected before publication, and the finding survived, weaker and truer.

This is why the raw generations ship with the dataset. Anyone who disagrees with how a trial was scored can rescore it from the text.

Licence

• Data: CC BY-SA 4.0, inherited from FLORES-200.
• Machine-translated payload strings additionally carry a non-commercial term from NLLB-200. Use them accordingly.
• The analysis code in the companion repository is MIT.

Citation

@misc{oliveira2026ragsec,
  author       = {Welerson dos Anjos de Oliveira},
  title        = {Does Language Resource Level Change Prompt Injection
                  Success in Offline {RAG}?},
  year         = {2026},
  howpublished = {UNIGOU Proceedings, Institute of Czech-Brazilian
                  Academic Cooperation (INCBAC)},
  url          = {https://www.incbac.org/unigou-proceedings/?article_id=919},
  doi          = {10.5281/zenodo.22861168}
}

Paper: https://www.incbac.org/unigou-proceedings/?article_id=919 Code: https://github.com/welerson-py/RAGSEC Archived deposit: https://doi.org/10.5281/zenodo.22861168