From 210fdb72ee63a134d8617e6e371097900881bc0d Mon Sep 17 00:00:00 2001 From: Demetris Gerogiannis Date: Thu, 9 Jul 2026 09:59:45 +0300 Subject: [PATCH] docs: remove unsourced performance claims; qualitative architecture-grounded language only Marketing flagged the 'sub-millisecond latency for most requests' tagline as contradicting our own tier table. Rather than swapping one unsourced number for another, all quantitative performance claims (per-tier timings, ~85% catch rate, ~10-15% escalation share) are replaced with qualitative, architecture-grounded statements until published benchmarks exist. Applies to the README tagline, tier diagram, and package summary. Co-Authored-By: Claude Fable 5 --- CHANGELOG.md | 10 ++++++++-- README.md | 14 +++++++------- pyproject.toml | 2 +- 3 files changed, 16 insertions(+), 10 deletions(-) diff --git a/CHANGELOG.md b/CHANGELOG.md index 6b69b61..3a0c49c 100644 --- a/CHANGELOG.md +++ b/CHANGELOG.md @@ -28,9 +28,15 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0 locally and free; provider env vars only needed for the Tier 3 LLM Judge), and the CLI section is reframed as "Train guardrails from your test results". Contributing/CLA wording cleaned up. +- **Removed unsourced performance claims** from the README tagline, tier + diagram, and package summary ("sub-millisecond latency", per-tier timings, + "~85% of prompt injections", "~10-15% of traffic"). Until we publish + benchmarks, performance language is qualitative and architecture-grounded + only: local tiers are free with no LLM call; the Tier 3 judge is the only + tier with token cost and is invoked only on escalation. - Package summary now reads "Multi-tier firewall for AI agents — blocks prompt - injections, jailbreaks, and scope violations with sub-millisecond latency - for most requests." + injections, jailbreaks, and scope violations; fast local tiers screen every + request, only uncertain cases reach an LLM judge." ## [0.2.1] — 2026-05-12 diff --git a/README.md b/README.md index 1942656..6da6ded 100644 --- a/README.md +++ b/README.md @@ -5,7 +5,7 @@

humanbound-firewall

- Multi-tier firewall for AI agents. Blocks prompt injections, jailbreaks, and scope violations with sub-millisecond latency for most requests. + Multi-tier firewall for AI agents. Blocks prompt injections, jailbreaks, and scope violations — fast local tiers screen every request; only the uncertain ones reach an LLM judge.
4-tier architecture · pluggable models · guardrails trained from your own test data

@@ -45,22 +45,22 @@ Every user message passes through four tiers before reaching your agent: ``` User Input | -[ Tier 0 ] Sanitization ~0ms, free +[ Tier 0 ] Sanitization no model call, free | Strips invisible control characters, zero-width joiners, bidi overrides. | -[ Tier 1 ] Basic Attack Detection ~15-50ms, free +[ Tier 1 ] Basic Attack Detection local model inference, free | Pre-trained models (DeBERTa, Azure Content Safety, Lakera, etc.) | Pluggable ensemble — add models or APIs, configure consensus. - | Catches ~85% of prompt injections out of the box. + | Catches the bulk of generic prompt injections out of the box. | -[ Tier 2 ] Agent-Specific Classification ~10ms, free +[ Tier 2 ] Agent-Specific Classification local model inference, free | Trained on YOUR agent's adversarial test logs and QA data. | Catches attacks Tier 1 misses. Fast-tracks legitimate requests. | You provide the model — we provide the training orchestrator. | -[ Tier 3 ] LLM Judge ~1-2s, token cost +[ Tier 3 ] LLM Judge LLM call, token cost Deep contextual analysis against your agent's security policy. - Only called when Tiers 1-2 are uncertain (~10-15% of traffic). + Only called when Tiers 1-2 are uncertain — a small fraction of traffic. ``` Each tier either makes a confident decision or escalates. No forced decisions. diff --git a/pyproject.toml b/pyproject.toml index 7944fd4..84d06ce 100644 --- a/pyproject.toml +++ b/pyproject.toml @@ -5,7 +5,7 @@ build-backend = "setuptools.build_meta" [project] name = "humanbound-firewall" version = "0.2.2" -description = "Multi-tier firewall for AI agents — blocks prompt injections, jailbreaks, and scope violations with sub-millisecond latency for most requests" +description = "Multi-tier firewall for AI agents — blocks prompt injections, jailbreaks, and scope violations; fast local tiers screen every request, only uncertain cases reach an LLM judge" readme = "README.md" license = "Apache-2.0" requires-python = ">=3.10"