NEAR AI has launched two private, “uncensored” artificial intelligence models aimed specifically at cybersecurity research, giving security teams access to models whose refusal behavior has been deliberately removed. According to the official NEAR AI cybersecurity page, DeepSeek V4.1 Flash 1017 Abliterated and Qwen 3.8 27B Abliterated are now available through NEAR AI Cloud for vulnerability analysis and authorized security work. The company announced the cybersecurity initiative on October 29.
NEAR AI describes the models as “abliterated,” meaning safety-related refusal behavior has been removed so they can respond directly to questions about weaknesses, attack surfaces and exploitation techniques that conventional commercial models may decline. The designation refers to model behavior rather than a separate cybersecurity protocol or scanning engine. Both models are open to every NEAR AI Cloud account without a separate application or mandatory identity check, according to the company.
Secure Enclaves Protect Sensitive Security Research
The cybersecurity models run under NEAR AI’s Private TEE tier rather than its less restrictive Incognito routing tier. Supported workloads execute inside hardware-isolated environments using Intel TDX and NVIDIA confidential-computing technology. Prompts are decrypted for inference inside the trusted execution environment, while NEAR AI says infrastructure operators cannot access the plaintext data processed inside that enclave. Hardware attestation can also be used to verify the environment handling a request.
That architecture builds on the confidential inference stack NEAR AI has been expanding, including hardware-backed attestation and isolated GPU workloads. The same privacy infrastructure has also been integrated with development environments through NEAR AI’s OpenCode support, allowing developers to route code-related inference through compatible confidential models. For cybersecurity teams, the practical objective is to analyze proprietary source code without exposing vulnerabilities to the infrastructure operator during inference.
The protection is not equivalent to eliminating every trust assumption. NEAR AI’s own terms acknowledge that TEEs can still depend on hardware manufacturers, firmware, microcode and attestation infrastructure and may remain exposed to theoretical side-channel or hardware-level vulnerabilities. Confidential computing therefore moves the trust boundary into protected hardware rather than making sensitive computation universally trustless.
Pricing also varies substantially between the two models. Qwen 3.8 27B Abliterated costs $0.04 per million input tokens and $1.35 per million output tokens, with a 262,000-token context window. DeepSeek V4.1 Flash 1017 Abliterated costs $0.30 per million input tokens and $1.20 per million output tokens while supporting approximately 1.05 million tokens of context. The advertised $0.04 starting price therefore applies only to Qwen input tokens, not to all inference generated by the cybersecurity models.
NEAR AI RED Targets Security Teams
Alongside the model release, NEAR AI introduced NEAR AI RED, a program aimed at red teams, security auditors and internal application-security groups. Membership includes 20% off the first month of private uncensored inference, early access to additional models, security workshops and access to a community of practitioners. Joining RED is optional and is not required to use the underlying cybersecurity models.
Participants in RED are required to use the models for lawful research or authorized testing on systems they own or have permission to examine. The distinction is relevant because removing model refusals increases the technology’s dual-use capability: the same reasoning that can identify vulnerabilities during defensive auditing can also describe offensive techniques. NEAR AI is separating unrestricted model responses from authorization to attack third-party systems, which remains governed by legal and program-use requirements.
The launch extends NEAR’s broader effort to combine AI infrastructure with privacy technologies, which has also included confidential payment infrastructure for AI agents. In cybersecurity, however, the immediate product is narrower and measurable: two live abliterated models, confidential execution for NEAR-hosted inference and an optional security-team program. The key differentiation is not simply fewer AI guardrails, but pairing that model behavior with an attested environment intended to keep sensitive code and vulnerability findings outside the cloud operator’s visibility.
Natalie Pierce tracks the parts of crypto that move fast and rarely wait for everyone to catch up. From South Africa, she covers DeFi, AI crypto, hacks, airdrops, sentiment and emerging narratives, especially when user behavior and protocol risk start to overlap.
Her reporting is built for messy sectors. Natalie looks at incentives, reactions, security concerns, social momentum and early signs of traction without pretending every new trend is already proven. Her voice is clear and accessible, but careful enough for areas where excitement can outrun the facts very quickly.