NEAR AI RED

NEAR AI RED is a cybersecurity program from NEAR AI (the AI arm of the NEAR ecosystem), launched on October 29, 2026 together with private, uncensored ("abliterated") models for security research. The models — abliterated versions of DeepSeek V4.1 Flash and Qwen3.8 27B — have their refusal behaviour reduced so they answer vulnerability questions instead of refusing, and they run inside trusted execution environments (TEEs) on NEAR AI Cloud so that, NEAR says, not even NEAR AI can see the prompts.

The models are open to every NEAR AI Cloud account; NEAR AI RED is the optional membership for red teams, auditors and AppSec teams, offering discounts, early access, workshops and support. On October 30–31 NEAR also released the open-source NEAR AI RED Sandbox, a disposable microVM for running security agents safely on Apple-silicon Macs.

What's included

Models (open to all NEAR AI Cloud accounts)

Model	Context	Price per 1M tokens (in / out)
DeepSeek V4.1 Flash (abliterated)	~1.05M	$0.30 / $1.20
Qwen3.8 27B (abliterated)	262K	$0.04 / $1.35

• Confidential inference in TEEs, with optional end-to-end encryption.
• Verifiable protections — hardware attestation and, where supported, signed responses.
• OpenAI-compatible API — switching is a base-URL change.

NEAR AI RED membership (optional, by application)

• 20% off the first month of inference
• Early access to new models and security features
• Private Slack channel and dedicated support
• Workshops, co-marketing and input into the roadmap

Members commit to use the models only for lawful security research or authorized testing on systems they own or have written permission to test.

RED Sandbox (open source) — each session gets a fresh microVM; the agent's commands and file edits run only inside it, files are copied in, network is blocked unless specific sites are allowed, the API key stays outside, and only files saved to the output folder come out. Early version supports Apple-silicon Macs only.

Why it matters

Mainstream models often refuse to help with exploit reproduction because they cannot tell defenders from attackers. NEAR's pitch is to give defenders the same uncensored capability attackers already use, but privately and verifiably.

1. Create a NEAR AI Cloud account and add balance; the abliterated models are available without applying.
2. Point any OpenAI-compatible client at NEAR AI Cloud and pick a model such as deepseek-ai/DeepSeek-V4.1-Flash-1017-abliterated or Qwen/Qwen3.8-27B-abliterated.
3. Optionally verify the TEE attestation before sending sensitive code.
4. To run security agents safely on a Mac, install the open-source nearai-red-sandbox from GitHub; each session runs in a fresh, network-isolated microVM.
5. Security teams that want discounts, early access and support can apply at near.ai/cybersecurity/apply.

Important: only test systems you own or are explicitly authorized to test, and have a human review any findings — uncensored models can produce harmful or wrong output.