Wired reported on 14 October that Tech Against Terrorism, a UK nonprofit, tested 134 AI models with over 200 requests a terrorist might make. Wired said 81 models failed and two passed fully. A model failed if it fully answered even one mass-casualty request, or scored under 90 overall. Claiming to be a safety researcher, with no proof, lifted the answer rate to 16.9 percent, almost ninefold.
Alibaba released Qwen 3.8-27B on 14 August. It is open-weight, so anyone can download it; closed-weight models stay with their makers. Within 24 hours, copies advertised as abliterated, meaning guardrails removed, appeared online. Wired did not say anyone checked that claim. Tech Against Terrorism says free tools can strip safety controls from openly released models. Its director, Adam Hadley, said guardrails can, as far as he knows, be removed completely. The sources do not say whether the 134 tested models included altered copies.
Once a model is downloaded, the user holds its files and can edit them. A maker's safety results describe the model as released, and the sources do not show they carry over to a changed copy. The tested models answered far more often after one unproven claim. Buyers can ask where a copy came from and who tested it.