Cracken Launches Red v3, Unlocking Offensive Cyber Capabilities With 7× Fewer Safety Failures
Cracken launches Red v3, built on a proprietary technique that unlocks attacker-grade cyber capabilities while keeping loss-of-control risk at the original model’s level.
SAN FRANCISCO, November 18, 2026. Cracken, an applied AI lab for proactive cybersecurity, today launched Red v3, the latest version of its offensive-security AI. Red v3 is built with Cracken’s proprietary cyber-specific abliteration technique, which removes only cybersecurity refusals from open-weight models instead of stripping safeguards across all categories. In testing, broadly uncensored alternatives violated operating constraints up to seven times as often as Cracken’s version, which stayed in line with the original model.
Offensive cyber capabilities are spreading quickly to open-weight AI. The U.S. Center for AI Standards and Innovation and Anthropic have both flagged the trend, and Anthropic reported that uncensored versions of a leading cyber-capable open-weight model appeared within days of its release.
This leaves enterprise security teams with a dilemma. Safety-aligned models refuse legitimate security work: on RedLineBench, Cracken’s open-source benchmark of 153 offensive-security tasks, widely used open-weight models refused up to 39% of requests. Closed-weight cyber models such as OpenAI’s Daybreak Red lack independent safety measurements, are served only through the vendor’s API, and cannot be deployed on-premises or in sovereign, air-gapped environments to meet enterprise data requirements. Broadly uncensored open-weight models remove those barriers, but also weaken the safeguards that prevent AI agents from violating operating constraints or being hijacked through prompt injection.
The difference is how AI is uncensored
Cracken compared four versions of Alibaba’s Qwen3.8-27B: the original, Cracken’s cyber-specific version, and two broadly uncensored versions publicly available on Hugging Face, one created with the open-source tool Heretic and one published by Huihui AI.
Loss of control: On ODCV-Bench, a public benchmark from McGill University researchers that tests whether AI agents break operating rules under performance pressure, Huihui’s version violated its constraints in 50% of scenarios and Heretic’s in 19.7%. Cracken’s version did so in 6.5% and the original in 7.7%. An independent AI judge supported the findings.
Prompt injection (preliminary): Malicious instructions hijacked Huihui’s version in 48.9% of attempts and Heretic’s in 23.3%, compared with 15.9% for Cracken’s version and 12.2% for the original.
Other safeguards: Cracken’s version still refused 98 of 100 generic harmful requests.
The technique carries across model generations. In Cracken’s research covering 24 open-weight models, it cut cybersecurity refusals on the trillion-parameter Kimi K2 from 100% to 7%, while other harm categories kept 44-100% of their refusals. Red v3 applies it to the GLM-5.3 model family.
“Attackers can already remove restrictions from the most powerful open-weight models. The unsolved dilemma was to make these models safe for defenders to use within their organizations for proactive cybersecurity,” said Artem Sorokin, Co-founder and CEO of Cracken. “Our approach gives security teams attacker-grade AI without dismantling its safeguards, and we can apply it to each more powerful model as it emerges.”
Red v3 is available now through Cracken’s proactive enterprise platform. Every engagement is scope-locked, non-reversible actions require human approval, and every action is logged. The platform supports cloud, on-premises and fully air-gapped deployments.

