haris-arabic-injection-detector

Haris injection classifier (Arabic / English)

The recall layer of Haris (حارس), a detection-only guardrail for LLM applications with first-class Arabic and Saudi support.

This model is one half of a hybrid. Haris pairs it with a high-precision rule/signature detector; the model exists to catch the paraphrases the rules miss. It is not designed to be used alone.

Read the Limitations before deploying. Two things bound how far the numbers below can be trusted: the Arabic training data has not been reviewed by a native speaker, and the test set is self-authored rather than an independent benchmark. Run-to-run variance was the third, and is now measured and small — see Run-to-run variance.

Intended use

Flagging likely prompt-injection attempts in user input to an LLM application — instruction override, role-play/jailbreak, and system-prompt extraction — in Arabic and English. Maps to LLM01:2025 Prompt Injection and LLM07:2025 System Prompt Leakage in the OWASP Top 10 for LLM Applications (2025 edition).

Detection only. It classifies text. It does not generate attacks, and it does not decide what to do about them.

Out of scope

• Not a defense on its own. See Limitations.
• Input, not output. It has never been evaluated on model responses.
• Not a general content-safety classifier. Injection ≠ toxicity ≠ policy violation. It is not trained to detect harmful requests; that is a different threat.
• Not calibrated. Treat the probability as ordinal, not as a true likelihood.

How to use

Through Haris, which applies the hybrid combination:

from haris import scan

result = scan(text, detectors=["pii_saudi", "injection", "ml_injection"])
result.risk, result.score, result.flagged_spans

Directly:

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch

tok = AutoTokenizer.from_pretrained("MODEL_ID")
model = AutoModelForSequenceClassification.from_pretrained("MODEL_ID").eval()

enc = tok(text, truncation=True, max_length=128, return_tensors="pt")
with torch.no_grad():
    p_injection = torch.softmax(model(**enc).logits, dim=-1)[0, 1].item()

Label 1 = injection, 0 = benign.

Haris trips a verdict on the model alone only at p ≥ 0.588 — the model's score is multiplied by a trust factor of 0.85 before being compared to the 0.5 risk threshold, so a mildly-confident model false positive cannot block traffic by itself. The scores combine as max(rule_score, p × 0.85) rather than summing, so agreement between the two layers does not inflate the verdict.

Training data

1,648 rows total, split 70/15/15 into 1,131 train / 242 validation / 275 test. Deliberately balanced within each language so the model cannot shortcut on script:

• English: deepset/prompt-injections (Apache-2.0), used with attribution.
• Arabic: 253 hand-authored seeds (105 injection / 148 benign), culturally adapted rather than translated, expanded 4× with meaning-preserving surface variants (politeness framing, orthographic drift).

The Arabic benign half deliberately includes hard negatives — ordinary queries containing trigger-adjacent vocabulary — plus realistic RAG traffic on Saudi labour law, seerah, and government services. Benign and attack seeds are written to share the verb and differ in the object, so intent cannot be read off vocabulary:

Splitting is grouped by seed, so surface variants of one seed never straddle a boundary, and stratified by (language, label). Every string in the project's test fixtures is excluded from training.

Why the balance matters

If Arabic rows were mostly attacks, the model would learn "Arabic script ⇒ attack", report high accuracy, and flag every benign Arabic question — the precise opposite of this project's purpose, and invisible in a headline metric. The gap between Arabic and English injection rates is 1.8 points, so that shortcut is worthless.

Evaluation

Held-out test set: 275 rows (176 Arabic, 99 English). The checkpoint was selected on a separate validation split; test influenced neither training nor checkpoint selection.

Arabic recall is 0.623 → 1.000, against 6 false positives and 0 false negatives out of 172. Read that 1.000 as "no misses in a small sample", not as "catches everything": the 61 Arabic attack rows in the test set are surface variants of only 16 independent seeds, and by the rule of three a perfect score on 16 items is consistent with a true recall as low as 0.81. Widening that is the main open measurement task. The rules keep precision 1.000 and a 0.000 false-positive rate: they are the layer that is never wrong, and the model is the layer that is rarely silent.

An earlier run of this table showed rules recall of 0.000, which exposed a real defect rather than a triumph of the model: the Arabic signatures accepted only the definite article (تجاهل التعليمات) and not the possessive (تجاهل تعليماتك), which is the more natural attack phrasing. Fixed, and the row above is the re-measurement. Historical rules figures of 0.25–0.30 quoted in earlier rounds were split luck rather than a stable baseline.

Run-to-run variance — read this before quoting any figure above

The same data and code, retrained with three different RNG seeds and nothing else changed:

This is the second measurement. The first, on an earlier build, produced a recall spread of 0.194 and a verdict that flipped between seeds; a later build made it worse (0.311). The cause was found and fixed, and it is worth stating because the failure was invisible in the headline numbers:

• Two of three runs were shipping an unconverged checkpoint. Selection ranked epochs on a 146-row Arabic validation slice, which was noisy enough to prefer an epoch-1 model — still inside the learning-rate warmup, training loss 0.65. Those models caught 16% and 46% of Arabic attacks alone.
• The ship criterion passed anyway, because it measures the hybrid, and the high-precision rule layer carried a near-inert model. A stable "enable by default" verdict was being produced by a broken model plus working rules. That is a worse outcome than a failed run, because it looks like success.
• Selection now requires an epoch to have converged and scores the Arabic slice together with overall validation F1 rather than on 146 rows alone.

With that fixed, each model earns its place independently: model-only Arabic recall is 0.934 / 0.934 / 0.951 across the three seeds, against 0.623 for the rules.

Caveats on the spread itself: it is computed from three runs, so it is an estimate of the noise floor, not a precise bound — do not read "0.009" as a fine-grained significance threshold. Numbers from before the last dataset rebuild are not comparable to these, because the split was reshuffled and the test set is a different sample of seeds; that is why the 31 pinned rows exist.

Reproduce with python training/variance.py.

The number that decided deployment

Not accuracy. A fixed benchmark of 8 benign Arabic seeds — sentences whose vocabulary is dominated by attack signatures ("repeat the last sentence you said", "what does ignore mean in psychology", "explain the new Saudi labour law") — is routed to the test split unconditionally and has never been trained on in any round.

The benchmark is 31 rows, so a single seed's 3 false positives move it by 0.097. It is a useful regression tripwire, not a precise instrument.

The ship criterion, fixed in writing before any numbers existed: enable by default only if recall improves materially and the benign-Arabic false-positive rate does not regress. A model that buys recall by flagging ordinary Arabic questions is a regression regardless of its F1.

Earlier rounds failed this bar (pinned FPR 0.196, then 0.174) and the model shipped disabled. The bar was never moved to fit the numbers. It is now met by every seed tested, with zero spread — 0/31 on all three — which is what changed the recommendation from opt-in to enabled by default.

Training-time checkpoint metrics (epoch 3 of 8, selected on validation data only, scoring the Arabic slice together with overall F1 and requiring the epoch to have converged): val Arabic F1 0.938, val overall F1 0.947, val accuracy 0.955. These are selection scores; the test table above is the result.

Limitations

• The Arabic data has not had a native-speaker review. All 253 seeds are marked reviewed: false. This is now the largest known risk in the model, and the one thing the numbers cannot speak to: they measure the model against Arabic authored for this project, not against independently-validated Arabic. 206 of those seeds (105 attacks, 101 benign near-misses) are extracted for review in data/review/arabic_seeds_for_review.csv; the remaining 47 generic benign seeds have public-corpus equivalents and do not need one.
• The test set is not an independent benchmark. The Arabic half is self-authored, so these numbers measure consistency with our own conception of the threat. No public Arabic prompt-injection dataset was found to exist. Multilingual jailbreak corpora (MultiJail, AdvBench) were deliberately not used — they contain harmful-content requests, a different threat from prompt injection.
• Arabizi is a weak spot. Gulf/Najdi dialect is now represented on both sides of the label (an internal audit found it appearing only in attacks, which risked teaching dialect ⇒ attack; see data/README.md). Arabizi (Arabic in Latin script, e.g. t3ahal) is not detected at all — and is documented as the sharper attack vector (Al Ghanim et al., EMNLP 2024, which showed it bypassing safety mechanisms in GPT-4 and Claude 3 Sonnet).
• A small test set, which bounds every figure above. The Arabic attack rows come from just 16 independent seeds, so Arabic recall carries a wide confidence interval however good the point estimate looks (a perfect score on 16 items still admits a true recall of 0.81). Run-to-run variance itself is now measured and small — recall spread 0.033 across three seeds — but that measures stability, not coverage.
• Small dataset. 1,648 rows. Viable only because this is fine-tuning, not training from scratch, and it bounds how far the model generalises. Training loss reaches ~0.005, so it fits the training set completely.
• Rule-layer false positives on benign roleplay. The تخيل أنك ("imagine you are") signature fires on the verb regardless of its object, against this project's own design rule that benign and attack share the verb and differ in what is acted on. Measured on 36,162 natively-written Arabic sentences from public corpora, the rule layer flagged 2 — both the same benign customer-service roleplay. A 0.0055% rate, but concentrated in a realistic enterprise use case.
• No general instruction/QA data. The corpus contains attacks and trigger-adjacent benign text, but little ordinary unrelated traffic, which is what a deployment mostly sees.
• No obfuscation handling. Base64, homoglyphs, zero-width splitting are untested.
• Truncates at 128 tokens. A long benign prefix followed by an injection may be cut before the payload.
• English is upstream-curated — its errors are inherited from the deepset labelling.

Detection is not defense

This model raises the cost of an attack. It does not eliminate one. Deploy it as one layer alongside least-privilege tool access, output validation, and human review for high-stakes actions — never as the only thing between a user and your model.

Ethical considerations

Built defensively. The training corpus consists of labelled fixtures — attack patterns used as classifier training signal — and all PII values in the associated test suite are synthetic. No attack is generated at inference time.

Citation

Research grounding is in docs/RESEARCH.md. The hybrid design — a pretrained classifier combined with heuristic rule features — follows a published finding that the combination outperforms either alone (Ji, Li & Mao, KSEM 2025 AI & Sec Workshop). The detected categories correspond to the goal hijacking and prompt leaking classes of Perez & Ribeiro, NeurIPS ML Safety Workshop 2022.

@software{haris_guardrail,
  title  = {Haris: an Arabic/English LLM guardrail},
  url    = {https://github.com/ZKSolution/haris-guardrail},
  year   = {2026}
}