⚠  MOCKUP ONLY — This site is not yet functional. All data shown is illustrative and not real.
ExtremelyPublic.ai

Independent AI Safety Research

Before AI reaches your hospital, your court, your school — is it safe?

AI models are being deployed at unprecedented speed, with no agreed safety standard and no independent oversight. We test every freely available model and publish every result — unfiltered, unsponsored, and fully transparent.

See the safety results ↓

Four dimensions of safety — every model, every time

Refuses to produce harmful content

Does the model generate dangerous instructions, hate speech, or content that could cause real-world harm?

Treats all groups equally

Does the model show systematic bias against people based on gender, race, religion, or other characteristics?

Tells the truth, even when it's unpopular

Does the model make things up, agree with false claims to please users, or manipulate people?

Can't be used as a weapon

Can the model be prompted to assist with cyberattacks, chemical or biological weapons, nuclear or radiological threats, harmful chemistry, or lethal autonomous weapons?

Safety Ratings

6 models · 4 safety tests · updated 5 March 2025

A/B Good C Fair D/F Poor
1
Llama 3.1 70B Instruct Meta · 70B
Harmful outputs 91
Bias & fairness 80
Honesty 82
Weapons risk 83
B Good
2
Gemma 2 9B IT Google · 9B
Harmful outputs 88
Bias & fairness 79
Honesty 77
Weapons risk 80
B Good
3
Phi-3.5 Mini Instruct Microsoft · 3.8B
Harmful outputs 85
Bias & fairness 74
Honesty 70
Weapons risk 75
C Fair
4
Qwen2.5 7B Instruct Alibaba · 7B
Harmful outputs 78
Bias & fairness 71
Honesty 68
Weapons risk 75
C Fair
5
Mistral 7B Instruct v0.3 Mistral AI · 7B
Harmful outputs 72
Bias & fairness 65
Honesty 63
Weapons risk 72
D Poor
6
Falcon 40B Instruct TII · 40B
Harmful outputs 58
Bias & fairness 52
Honesty 49
Weapons risk 57
F Failing

Scores are 0–100 where 100 is perfectly safe. Grades: A ≥90 · B ≥80 · C ≥70 · D ≥60 · F <60. Each grade reflects the weighted average across all benchmarks in that category.

The idea

Like crash tests for cars.
But for AI.

Before a car reaches your driveway, it has been crash-tested by an independent body. The results are public. You can look up any model and see exactly how it performed. That transparency is why we trust the roads.

AI has no equivalent. Models are released every week, safety-tested only by their creators — if at all. ExtremelyPublic.ai is building that missing infrastructure: standardised, independent, and fully open safety ratings for every AI model anyone can download.

01

Same tests, every time

Every model is run through an identical set of safety benchmarks built on Inspect AI, the UK AI Safety Institute's open evaluation framework.

02

Nothing hidden

All methodology, all scores, and all raw results are published openly. Anyone can verify, challenge, or reproduce our findings.

03

Continuously updated

As new models are released, community contributors run the evaluations and submit results. No central bottleneck.

Community-driven

Anyone can contribute a result.

New AI models appear every week. No single team can keep up. ExtremelyPublic.ai is designed so that any pre-verified contributor with a capable machine can run the safety suite on a new model and publish the results — in a single command.

The tool downloads the model, runs every benchmark, and submits a signed result to the leaderboard. No deep technical knowledge required.

# Evaluate any model in one command

$ extremelypublic run meta-llama/Llama-3.1-70B

→ Downloading model…

→ Running 104 safety benchmarks…

→ Publishing results to leaderboard…

✓ Done. Results live at extremelypublic.ai

CLI tool is in development. Join the mailing list to be notified at launch.