Amazon
Rank #19 of 93
Est.

Amazon Nova Premier: LLM Security & Jailbreak Resistance

As of 2026-07-04, Amazon Nova Premier has an attack success rate (ASR) of 20.0% on Guardion's LLM Vulnerability benchmark — ranking #19 of 93 models. It is moderately resistant — a meaningful share of adversarial prompts succeed.

Overall ASR
20.0%
lower is safer
Robustness
80%
100 − ASR
Rank
#19
of 93 models
Zero-Shot ASR
not measured

Attack Success Rate (ASR) is the share of adversarial prompts that elicit harmful output across zero-shot, TAP, and Crescendo attacks (HarmBench framework). Lower is safer. "Est." models are calibrated from public safety evaluations pending a Guardion benchmark run.

How Nova Premier compares

Amazon Nova Pro
Rank #24 · ASR 24.0% (est.)
Compare
xAI Grok-4 Heavy
Rank #18 · ASR 18.0% (est.)
Compare
OpenAI GPT OSS 20B
Rank #20 · ASR 20.0% (est.)
Compare
Google Gemini 2.5 Pro
Rank #17 · ASR 16.1%
Compare

Frequently asked questions

How secure is Amazon Nova Premier against jailbreaks?

Amazon Nova Premier is moderately resistant — a meaningful share of adversarial prompts succeed. On Guardion's LLM Vulnerability benchmark it records a 20.0% attack success rate (80% robustness), ranking #19 of 93 models (estimated from public safety evaluations).

What is Nova Premier's attack success rate (ASR)?

Nova Premier's overall ASR is 20.0% — the share of adversarial prompts that elicit harmful output across zero-shot, TAP, and Crescendo attacks. Lower is safer.

Is Nova Premier more secure than Nova Pro?

Nova Premier is more robust: 20.0% ASR vs Nova Pro's 24.0%. See the full head-to-head comparison for the per-attack breakdown.

How can I make Nova Premier safer to deploy?

Even robust models are bypassed under sustained attack. An inline runtime guardrail that classifies prompts and responses — like Guardion's prompt-defense models — blocks jailbreaks and prompt injection before they reach Nova Premier.

Harden Nova Premier in production

Guardion's runtime guardrails block jailbreaks and prompt injection before they reach any model — with sub-130ms policy decisions.

Request a demo