Energy per request (N7)
Measured on one shared server (AMD Ryzen 7 PRO 8700GE, 16 logical cores) with the CPU package energy counter (RAPL package-0, memory not covered). The draw of the other services on the box is removed with idle windows before and after every measurement window; an idle control and a positive control (one spinning thread) check that the method sees what it should.
| Measurement | n | Result |
|---|---|---|
| Idle control (idle window against its neighbours) | 6 windows | mean +0.28 W, sd 0.55 W — pass |
| Positive control (1 thread spinning, 6 s) | 3 rounds | +12.9 W / +13.5 W / +13.0 W — pass |
| AEGIS fast path, one scan | 49,806 scans | 5.7 mJ per scan (median of 3 rounds) |
Reference LLM request (qwen2.5:1.5b, 128 output tokens, CPU, 4 threads) | 12 requests | 691 J per request (rounds 434 J / 767 J / 691 J) |
| Energy not spent when a prompt is blocked (reference request minus one scan) | — | 691 J per blocked request |
The last row is per blocked request. How many requests a deployment blocks is not measured here, so no saving per day or per customer is derived from it. The reference model stands in for “the LLM call that does not happen”; a GPU or a hosted model costs a different amount per request.
Source: packages/aegis/bench/energy/results/n7-2026-10-29.json in the destill-trust repository (not public), commit bebd845 (FORTRESS #253), generated 2026-10-29. Method and limits: packages/aegis/bench/energy/README.md at the same commit.
Latency per AEGIS path (N7)
Wall clock around one scan, one process, deterministic mode, on the same server. Corpus: deepset injection and benign prompts (pinned), 563 prompts.
| Path | n | p50 | p95 | max |
|---|---|---|---|---|
| Fast path, first time a text is seen | 563 | 0.31 ms | 1.12 ms | 2.9 ms |
| Verdict cache (same text again) | 563 | 0.004 ms | 0.012 ms | 0.25 ms |
| Optional model tier (uncertain prompts only, waited for) | 21 | 5,007 ms | 17,431 ms | 56,425 ms |
21 of the 563 prompts fell into the uncertainty band that calls the model tier. That tier is off by default in the gateway; when switched on, it runs asynchronously or with a time budget, because on this server a waited-for model call would set the slow end of the distribution.
Source: packages/aegis/bench/energy/results/n7-2026-10-29.json, key latency_ms, destill-trust commit bebd845.
Next to three open-source filters: latency only (N8)
The same 5,158 prompts, in the same order, one request at a time, on one desktop machine (AMD Ryzen AI Max+ 395, 32 logical CPUs) that was also running other work. AEGIS runs single-threaded in Node; the three model filters run PyTorch on the CPU with 16 threads. All four passed their controls: five textbook injections flagged, five everyday requests passed.
| Filter | n | p50 | p95 | max | Controls (flagged / passed) |
|---|---|---|---|---|---|
| AEGIS (destill-trust) | 5,158 | 1.21 ms | 11.9 ms | 916 ms | 5 of 5 / 5 of 5 |
| ProtectAI DeBERTa v3 base prompt-injection v2 | 5,158 | 201 ms | 862 ms | 16,215 ms | 5 of 5 / 5 of 5 |
| LLM Guard PromptInjection scanner (same model weights, threshold 0.92) | 5,158 | 113 ms | 471 ms | 5,917 ms | 5 of 5 / 5 of 5 |
| Llama Prompt Guard 2 (86M) | 5,158 | 107 ms | 645 ms | 280,392 ms | 5 of 5 / 5 of 5 |
Read the latency columns as an order of magnitude on a loaded desktop, not as a ranking: ProtectAI and LLM Guard run the same weights and still took 1,644 s and 876 s for the whole run, a difference that most likely comes from other load on the machine, which was not recorded. Llama Prompt Guard 2 had 53 single requests over 10 s, clustered while other jobs saturated the machine; its maximum is that stall, not its speed. The model filters cut prompts at 512 tokens; AEGIS reads the whole text. AEGIS was called directly, without its verdict cache. Energy was measured for AEGIS only (N7, on the other machine), so no energy comparison is made.
Source: packages/aegis/bench/competitors/README.md, rendered from results/raw-aegis.json, raw-protectai.json, raw-llmguard.json, raw-promptguard2.json (all four over the same items, items_sha256 a14deb27…), destill-trust commits 992cd2a (FORTRESS #254, ProtectAI and LLM Guard, runs of 2026-10-30), 2f70072 and b669aba (FORTRESS #282, Llama Prompt Guard 2 and AEGIS on the current engine, runs of 2026-11-03).
Try it on your own traffic
A millisecond on our desktop says nothing about your prompts, your languages or your false-alarm budget. In a four-week pilot the gateway sits in front of your model on your own traffic, and the success measure (latency at your p95, false alarms you accept) is agreed in writing before it starts.
What is not on this page, and why
- Detection rates. The comparison file also counts how many attacks and harmless prompts each filter flagged. We do not publish those counts yet: the founder decided on 2026-10-25 that no detection figure goes public before the benchmark itself can be re-run by anyone from a public repository. The known gaps stay named: role-play jailbreaks, false alarms on non-English text, paraphrases.
- No “faster”, “leaner” or “better” claim. The tables above say what was measured, on which machine, with what n.
- No compression or quality score. The earlier NI-Stack SDK returned a fixed savings figure and a fixed quality score on every call. Its successor reports neither.
Using it from code
Any OpenAI-compatible client can talk to the gateway. The client SDK (packages/firewall-client in destill-trust, FORTRESS #256) adds one thing: a blocked prompt arrives as its own error type, so the calling code can tell “the filter said no” from “the model failed”. It is tested against the gateway code itself, started locally in the test.
import { FirewallClient, FirewallBlockedError } from '@destill/firewall-client';
const fw = new FirewallClient({ baseUrl: 'http://127.0.0.1:8787' });
try {
for await (const delta of fw.stream({ model: 'ollama/llama3.2:1b', messages: [{ role: 'user', content: question }] })) {
process.stdout.write(delta);
}
} catch (e) {
if (e instanceof FirewallBlockedError) console.log('blocked by AEGIS:', e.verdict); // no model call happened
else throw e;
}