AI NewsInfrastructureAnnouncement
Morph sells a hosted uncensored DeepSeek for AI red teaming
Morph is now selling DeepSeek V4.1 Flash with its refusals removed, at 500 tokens per second. Access is gated.

Why it mattersAttackers use models that never refuse. Defenders get refused on the same prompt. This product tries to close that gap.
A security team testing its own AI product is blocked most of the time. The model it is testing with refuses. The attacker trying to break the same product is blocked by nothing. Morph says this asymmetry is the real problem. Today the company posted its answer on Y Combinator.
Morph is now selling DeepSeek V4.1 Flash 1017 with its refusals removed. The service runs at 500 tokens per second, with a one million token context window. Access is gated to verified companies, Morph says. A business email and a stated use case are required. Every request is reviewed by hand.
The pricing and the limits
Morph lists the model as morph-dsv41flash-abliterated. The API is compatible with OpenAI clients, Morph says. Pricing is $0.30 per million input tokens, Morph says on the launch page. Cached input is $0.03 per million. Output is $1.20 per million. The company has not published a figure for latency beyond the headline 500 tokens per second.
The intended users are red teams, incident responders and malware analysts, Morph wrote in its 23 October blog post. Morph wrote that refusals block "2 to 3 of 5 runs" of legitimate security work on the mainstream hosted models. The 23 October post cites 6,644 uncensored models on Hugging Face. Morph argues that gating only the defenders to safe models leaves the attack side with every tool.
The paired service
Morph also sells a Persistent Red Teaming service alongside the model. It starts at $6,000 a month with unlimited token usage, Morph says. The company offers a refund if no high-severity vulnerability is found. Morph states that it "has gotten privileged access to sensitive information at 80% of the Series A+ startups". That figure is Morph's own, measured on its own customer sample. Morph has not published a method.
The service runs "swarms of security agents" against the customer's product, Morph says. The agents use the uncensored model. That way they do not refuse to try the attacks a real adversary would. The service page points to an external partner, MindFort, for some of the work.
What a buyer should notice
The practical question is simple. Is a team's current red team passing because the test model keeps refusing? If yes, the result is not a security signal. Morph's pitch is built on that problem. The hosted path also makes an uncensored model easier to pay for than to run.
The counter is that an uncensored model is useful to an attacker too. Morph's answer is manual access review. Whether that review holds up is the open question, and Morph has not described how it decides.
Source
Morph on Y Combinator and the company's own Why We Serve an Uncensored DeepSeek post by founder Tejas Bhakta.
This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.
Get AI News in your inbox
New developer tools, model and agent releases, and how teams are actually using them to release software. Short, and only when there is something worth reading.

/filters:no_upscale()/news/2026/11/uber-secure-ai-framework/en/resources/1uber-1-1794769459402.jpg)