Series BWe raised $41M to protect the internet from AI-powered abuseRead announcement

CINDER · 2026

Case study

How Cinder Helped Black Forest Labs Launch FLUX.2 Without Compromising on Safety

Black Forest Labs validated FLUX.2 against real-world adversarial misuse before launch, using Cinder to test the model across the highest-severity generative AI harms.

01

The brief

How Cinder helped Black Forest Labs launch FLUX.2 without compromising on safety

Black Forest Labs was preparing to launch FLUX.2, their latest family of generative AI models, to millions of developers. Black Forest Labs is committed to preventing adversarial misuse, including to generate child sexual abuse material and non-consensual intimate imagery.

Cinder red-teamed the model across four iterative cycles, observing >90% reduction in harmful outputs and 10x safer performance than industry peers. Evaluation turnaround was sub 48 hours.

The outcome: Black Forest Labs launched one of the most significant open-weight and closed-source model releases in generative visual AI with confidence their safety posture could withstand real-world adversarial misuse.

02

The challenge

Traditional red-teaming has a velocity problem.

The standard model for pre-launch safety testing is sequential: assemble a team, run prompts, produce a report, hold the launch. At the pace of frontier model development, that cycle is too slow and too brittle. By the time the report is written, the model weights have changed, the fine-tune has been updated, and the findings are already partially stale.

What Black Forest Labs needed wasn't a one-shot pre-launch audit with a sign-off attached. It was a testing partner that could keep pace with iterative development and deliver actionable adversarial findings while engineers were still in the codebase to act on them.

Cinder provided rigorous adversarial testing that matched our release velocity. Their team found important edge cases and helped us address them before launch.

Ben Brooks

Head of Public Policy, Black Forest Labs

03

The approach

Four iterative cycles. Two highest-severity harm categories.

Cinder evaluated both text-to-image and image-to-image generation across the two highest-severity harm categories for generative AI: CSAM and NCII. Four iterative evaluation cycles, each targeting different attack sophistication levels — including linguistic evasion, semantic manipulation, and multi-modal exploits.

Every unsafe output was categorized, every successful defense analyzed, and vulnerability patterns shared with Black Forest Labs' research team while the engineering team was still in the codebase. When FLUX.2 launched, Black Forest Labs had a documented, benchmarked posture — not an open question about what the model could be made to produce.

The pay off

Safer at launch. Verified, not assumed.

  1. >90%

    Reduction in CSAM and NCII vulnerability across text-to-image and image-to-image attack vectors.

  2. 10X

    Safer than benchmark industry models at launch.

  3. <48 HRS

    Evaluation turnaround per iterative red-team cycle.

04

Why Black Forest Labs Chose Cinder

Speed at enterprise quality. Cinder delivered comprehensive adversarial evaluations in under 48 hours, not weeks, without sacrificing rigor or attack sophistication.

Adversarial depth. Cinder's red team deployed advanced attack techniques across linguistic evasion, semantic manipulation, and multi-modal exploits that automated tools and ad hoc red teams routinely miss.

Quantifiable benchmarking. Identical prompt sets applied across Black Forest Labs and competitor models provided objective proof of improvement, not anecdotal claims.

Pre-launch, not post-mortem. Issues were identified and resolved before public release and not after a crisis.

05

What This Unlocked

Black Forest Labs shipped FLUX.2 knowing their safety posture was validated against real-world threats, not just internal testing.

Black Forest Labs developed a robust evidence base to support a public launch.

Their customers had the confidence to deploy at scale.

The ecosystem proved that open-weight models can be both safe and performant.

Cinder and Black Forest Labs continue expanding their partnership to new modalities and emerging attack vectors as generative AI evolves.

Learn more about Cinder's red team capabilities.

Ship with confidence

Talk with the team that helped Black Forest Labs launch FLUX.2 with a verified safety posture — adversarial testing that moves at your release velocity, not behind it.