Acceptable use policy.
Do not use Erebine to break the law, harm people, or push abuse through our systems. These rules are part of the Terms and apply to you and to anyone you build for.
// 01
Absolutely prohibited
- Child sexual abuse material (CSAM), whether real or AI-generated, and any sexualization or grooming of minors. We report CSAM to NCMEC and terminate the account. Zero tolerance.
- Non-consensual intimate imagery, including AI-generated forgeries.
- Content that is illegal to create or distribute, or that facilitates clear illegality: trafficking, terrorism, fraud, or weapons and CBRN procurement.
// 02
Not allowed
- Infringing intellectual-property, privacy, or publicity rights.
- Malware, spam, phishing, or credential harvesting.
- Harassment, threats, doxxing, or incitement to violence or self-harm.
- Probing, scanning, or fuzzing the router or workers outside documented endpoints; bypassing quotas, rate limits, isolation boundaries, or worker leases.
- Misrepresenting AI output as human-generated where that deceives or harms.
// 03
Models and uploads
- You must hold the rights to any model, adapter, dataset, or file you upload.
- You may not upload weights or data whose export or use is restricted by US law.
- We may scan, quarantine, or remove uploads to enforce this policy or the law.
// 04
GPU agent images
The GPU agent images carry vendor software under vendor terms. Those terms restrict what you may do with the image, they reach the whole image and not only the vendor libraries in it, and they bind you whether you run the hosted service or pull the image yourself. /docs/license names them and says where each license text sits inside the image.
On the NVIDIA (CUDA) agent images, the NVIDIA Software License Agreement and the Product-Specific Terms for NVIDIA AI Products prohibit:
- Emotion recognition. Prohibited outright, for the software and every component of it. Erebine ships image, video, and audio input, batch transcription, and realtime voice; none of it may be pointed at this.
- Offering a voice-cloning text-to-speech application as a service to third parties, or shipping one in your own product. Building one for your own use is permitted, and the content you feed it has to be content you hold the rights to.
- Running the NVIDIA SDKs, drivers, or Omniverse on anything but NVIDIA hardware, and replacing an NVIDIA component with other software that implements NVIDIA's APIs.
- Reverse engineering the NVIDIA binaries, or reverse engineering the output an NVIDIA SDK generates.
- Bypassing a technical limitation, encryption, security, rights-management, or authentication mechanism in the software.
- Using the software or NVIDIA confidential information to develop a competing product or technology, or to support an intellectual-property assertion against NVIDIA.
- Putting the NVIDIA components under an open-source or shareware license, and stating or implying that NVIDIA sponsors or endorses what you built.
- Combining paid and unpaid NVIDIA software in a way that avoids fees or exceeds a use limit.
NVIDIA does not test or certify this software for critical applications: avionics, navigation, autonomous vehicles, military, medical, life support, or any other system whose failure causes injury, death, or catastrophic damage. Deploy it there and the safety and redundancy engineering is yours. Use also has to stay consistent with NVIDIA's Trustworthy AI terms.
Benchmarks are allowed. You may publish benchmarking, regression, and performance results for the CUDA agent image without asking NVIDIA: the NVIDIA container benchmarking exemption names the cuda-dl-base image these are built on. The exemption does not name competitive analysis, which the agreement restricts in the same sentence, so get NVIDIA's written permission before publishing a comparison against a non-NVIDIA platform.
The AMD (ROCm) agent images carry no vendor use restriction of that kind. AMD licenses ROCm per component, under MIT, Apache-2.0, NCSA, and BSD licenses, and what those ask of you is notice: keep the copyright and license text with any copy you pass on. Two ROCm packages, ROCprof Trace Decoder and the AOCC CPU optimizations, ship in binary form under the AMD Software End User License Agreement instead, and using, copying, or distributing either one accepts that agreement.
These restrictions pass through. If you give anyone else access to an image, or to a product or service built on one, you have to bind them to terms at least as protective first. Section 18 of the Terms sets that out.
// 05
If you build for others
If you expose the service to your own users through your endpoints:
- You are responsible for your end users' use of the service.
- You must pass through equivalent acceptable-use restrictions to your users and moderate their use. We provide hooks to help.
// 06
Enforcement
We may investigate suspected violations and may throttle, suspend, or terminate accounts. Confirmed CSAM or sanctioned-party use results in immediate termination and, where required, a report to authorities. Appeals: contact@erebine.ai.
// 07
Reporting
- Report content or abuse: contact@erebine.ai or the in-product report control.
- Copyright: /dmca.
- Non-consensual intimate imagery: contact@erebine.ai (48-hour intake).