REDDIT · REDDIT// 3h agoBENCHMARK RESULT

Arc Gate claims perfect prompt-injection detection

Arc Gate is a prompt-injection filtering proxy for direct LLM traffic, positioned in front of any OpenAI-compatible endpoint. The launch claims perfect precision, recall, and F1 on a 40-prompt out-of-distribution benchmark covering indirect requests, roleplay, hypothetical scenarios, and technical framings, with zero false positives and zero misses. It reportedly blocks harmful prompts before they reach the model, adds roughly 350ms of detection overhead, and requires no GPU on the customer side.

// ANALYSIS

The core story is strong: if these numbers hold up outside a tiny benchmark, Arc Gate is solving a real deployment pain point with a low-friction architecture.

–The benchmark framing matters: 40 prompts is directional, not conclusive, so reproducibility and test-set disclosure are the key questions.
–The product’s practical appeal is the integration model: it sits in front of existing OpenAI-compatible endpoints instead of requiring model-side changes.
–The stated latency profile is acceptable for many safety-sensitive flows, especially if the proxy is blocking before upstream spend occurs.
–The comparison set is credible for this niche, and the recall gap versus Moderation and LlamaGuard is the most important claim here.
–The biggest risk is overfitting to a narrow attack distribution; indirect prompt injection systems tend to degrade when adversaries change tactics.

// TAGS

llm securityprompt injectionapi proxyopenai-compatibleai safetymoderationopen source

DISCOVERED

3h ago

2026-04-28

PUBLISHED

5h ago

2026-04-28

RELEVANCE

8/ 10

AUTHOR

Turbulent-Tap6723