PromptGuard vs open-source guardrails
The open-source guardrail libraries are good, genuinely free, and the right answer for a lot of projects. We would rather say that plainly than pretend otherwise. What they are not is a maintained detection service: you own the model updates, the evasion coverage, the latency budget and the on-call. The common end state in this market is layered, cheap deterministic checks in the application, a maintained firewall at the gateway.
We wrote this, so read it accordingly. Everything in the Open-source guardrails column was taken from their own public material, read 11 August 2026, and each claim links its source with the date we read it. Where their material does not state something we say that, rather than claiming they cannot do it. Our own figures come from the same constants the rest of this site renders, and what we say about how PromptGuard works follows our own documentation.
- Cost
- PromptGuard
Public. $0 with $5 of usage a month and no card, then $20/month.
Open-source guardrailsFree, and self-hosted by definition. You pay in engineering time and compute.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
- Who maintains detection quality
- PromptGuard
We do, and we publish the benchmark each time it moves.
Open-source guardrailsYou do, together with the project's maintainers and release cadence.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
- Evidence behind the detection claim
- PromptGuard
F1 0.887, 95% CI [0.874, 0.900], across 2,369 samples from 8 datasets (5 independently published). Methodology and datasets named.
Open-source guardrailsVaries by project; community benchmarks exist but there is no single published figure across them.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
- Obfuscated and encoded attacks
- PromptGuard
Normalization runs ahead of detection, and we publish how the layered pipeline scores against a single classifier on our own evasion suite.
Open-source guardrailsDepends entirely on which scanners you enable and how you chain them.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
- Audit and compliance
- PromptGuard
Exportable audit trail, HMAC-SHA256 response signing, versioned policy with shadow-mode preview.
Open-source guardrailsWhatever you build around the library.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
- Runs inside your network
- PromptGuard
Managed cloud, hybrid self-host in your VPC, or air-gapped behind default-deny egress.
Open-source guardrailsYes, it runs wherever you run it. This is a genuine strength of the approach.
Source: LLM Guard, read 11 August 2026; NVIDIA NeMo Guardrails, read 11 August 2026
Choose Open-source guardrails when
- You are prototyping, or the application is internal and low-risk, and the cost of a miss is small.
- You have security engineers who will own detection quality, keep the rules current against new evasion techniques, and carry the on-call.
- You need to modify the detection logic itself, or licensing requires a permissive open-source component.
- Budget is the binding constraint. Free is a real advantage and we are not going to argue with it.
Choose PromptGuard when
- You want someone else maintaining detection as attacks change, with a benchmark published each time.
- You need the audit trail, tamper-evident logging and policy versioning that a compliance review asks for.
- You want obfuscated attacks covered without building the normalization layer yourself.
Sources: LLM Guard (read 11 August 2026) · NVIDIA NeMo Guardrails (read 11 August 2026) · Our benchmark methodology. Think we have something wrong? Tell us and we will correct it.
One engine in front of the AI you buildand the AI your people use
Your applications swap a base URL, no code change, and are protected in under 2 minutes. On your people's devices, an agent inspects the AI tools it covers and names the ones it does not. Every request is judged before it reaches the provider, in our cloud or entirely inside your network.