Injection guard

Perimeter prompt-injection defense combining fast regex filters and DeBERTa classification.

problem it solves
Blocks instruction overrides, role hijacks, and untrusted retrieval data from hijacking model control flow.

What it does

What it does: Filters malicious prompt injections and jailbreaks before prompts reach model execution.

What it watches: Known attack signatures, role override phrases, system prompt leakage, and vector embeddings.

When it triggers: When input patterns match regex firewall rules or classifier score exceeds threshold.

The Action: Rejects compromised requests with HTTP 400 or strips injection vectors in shadow mode.

How it recovers: Clean inputs pass directly to upstream model providers.

Regex Firewall Rulesrequired
Deterministic instruction-override patterns run at zero memory overhead.
Memory for DeBERTa Classifier (Optional)recommended
Learned stage needs ~2.2GB free RAM; falls back to regex-only firewall if constrained.

Mode	Effect on your request	What you can see
off	Not consulted. All prompts pass through without injection inspection.	No injection_guard stage recorded.
shadow	Scores prompts and records potential injection attempts without blocking requests.	Stage with action=would_block and risk_score logged.
prod	Refuses requests exceeding injection risk threshold with HTTP 400 Bad Request.	Stage with action=blocked and violation details logged.

Regex Stage	<0.1ms latency	Catches direct instruction overrides and exfiltration phrases.
DeBERTa Classifier	95.25% accuracy on benchmark	Catches paraphrased and retrieval-borne attacks.
Blocking threshold	Score ≥ 0.85	Configured to minimize false positive blocks.

·Security researchers or red-team prompts will trigger injection defenses.
·RAG retrieval context from untrusted websites can trigger indirect injection alerts.
·DeBERTa model loading adds ~2.2GB RAM requirement to gateway process.