terminal Live Prompt Input

Live Stream Active

Try:

Triggered Heuristic Rules

0 triggered

Rule ID	Pattern Description	Risk
Run a scan to see triggered rules

Combined Risk help

— Awaiting

Model Prob

—

TF-IDF

—

Heuristics

— radio_button_unchecked

Entropy

— analytics

Complexity

— segment

Flagged Rules

— list_alt

psychology Contextual Analysis

Using DistilBERT fine-tuned on the deepset/prompt-injections dataset with 8 regex heuristic rules and TF-IDF cosine similarity.

DistilBERT v1 8 Heuristics TF-IDF

Weight Tuning Simulator

Calibrate the hybrid classifier's decision engine. Adjust weights to explore how classification changes. Go to Inspector first to get live scores.

Weight Configuration

Total: 1.00

history_edu w_heur (Heuristic) 0.40

Regex pattern matching — fast & deterministic

psychology w_model (Neural) 0.40

DistilBERT semantic intent analysis

analytics w_tfidf (Statistical) 0.20

TF-IDF cosine similarity to known injections

Decision Threshold

0.50

SENSITIVE STRICT

Score >= threshold → injection blocked. Currently: balanced detection.

0.42 Total Score

check_circle SAFE STATE

functions

Score Formula


            Score = (0.40 × S_heur) +
            (0.40 × S_model) +
            (0.20 × S_tfidf)

S_HEUR —

S_MODEL —

S_TFIDF —

Batch Dataset Analysis

Validate bulk LLM inputs against injection patterns. Paste prompts (one per line) below.

0 prompts

Results — Flagged Samples

ID	Prompt Snippet	Primary Rules	Score	Decision

Diagnostic Metrics Center

Real metrics from your trained model on the held-out test set and red-team suite.

ACCURACY

—

PRECISION

—

RECALL

—

ROC AUC

—

grid_view Confusion Matrix

TRUE POSITIVE —

FALSE POSITIVE —

FALSE NEGATIVE —

TRUE NEGATIVE —

Predicted SafePredicted Threat

Red-Team Attack Suite

Total Prompts

—

Blocked

—

Block Rate

—

visibility False Positive Inspector

Loading false positive data...

AegisPrompt