arc-sentry 3.1.0

Arc Sentry — prompt injection detection for LLMs. DR=100%, FPR=0% on Mistral-7B. Geometric detection via Fisher manifold (Nine 2026).

Arc Sentry v3.0.0

Pre-generation prompt injection detection for open source LLMs.
Blocks attacks before model.generate() is called.

---

Benchmark — v3.0.0

Metric	Result
Detection rate	100%
False positive rate	0%
Session requests	450
Latency	42ms/req
Layer SNR (Mistral 7B)	2.053
FR separation	0.0787

450-request session benchmark on Mistral-7B-Instruct-v0.2.
270 normal requests, 180 injection attempts (dense + subtle roleplay/hypothetical).
Zero false positives across all safe blocks.

Also validated: Garak promptinject suite 192/192 blocked, Crescendo flagged Turn 3 (LLM Guard: 0/8).

---

Install

pip install arc-sentry

Usage

from arc_sentry import ArcSentryV3, MistralAdapter
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

model = AutoModelForCausalLM.from_pretrained(
    "mistralai/Mistral-7B-Instruct-v0.2",
    torch_dtype=torch.float16, device_map='auto')
tokenizer = AutoTokenizer.from_pretrained('mistralai/Mistral-7B-Instruct-v0.2')

adapter = MistralAdapter(model, tokenizer)
sentry = ArcSentryV3(adapter, route_id="my-deployment")
sentry.calibrate(warmup_prompts)  # ~100 prompts from your deployment

response, result = sentry.observe_and_block(user_prompt)
if result["blocked"]:
    pass  # model.generate() was never called

How it works

Three detection layers:

1. Phrase check — 80+ injection patterns, zero latency
2. Geometric detection — mean-pooled hidden states at optimal layer, Fisher-Rao distance from calibrated centroid. Catches injections with no explicit language.
3. Session D(t) monitor — stability scalar over rolling request history. Catches gradual campaigns (Crescendo-style) invisible to single-request detection.

Grounded in the second-order Fisher manifold (H2 x H2, R = -4, tau* = sqrt(3/2) ~= 1.2247).
Full theory: bendexgeometry.com/theory

Detection mechanism

1. Mean-pool hidden states at layer L (validated: L=16 on Mistral-7B)
2. L2-normalize: h = h / ||h||
3. Fisher-Rao distance to warmup centroid
4. Distance > threshold -> BLOCK (phrase check runs in parallel)
   model.generate() is never called

Also available

• Arc Vigil — training stability monitor. 100% detection, 0% FP, 90% auto-recovery.
  pip install arc-vigil

---

Bendex Geometry LLC · Patent Pending · 2026 Hannah Nine
bendexgeometry.com · PyPI