Verified, then logged.
Sun 22 Nov 2026 · 00:21 KST / 15:21 UTC한국어 | ENSearch /

Breaking04 / Policy

Korea AI Safety Institute withholds uncensored model results

The institute published a SHA-256 hash in place of the report, leaving the differences in harmful responses between original and modified models undisclosed.

Two identical workstations sit side by side on a long bench in a research institute's evaluation room, their screens turned away from the camera, as a blurred researcher stands with their back turned at the far end of the room. AI-generated image

Korea's AI Safety Institute (AISI), the body responsible for the country's cooperation with AI safety institutes abroad, announced on 3 Nov 2026 that it had evaluated open-weight AI models stripped of their safety alignment for harmful output. The notice discloses that the evaluation took place and gives a hash of the report, but does not publish the detailed results.

According to the notice, the evaluation covered GLM 5.3 Flash and Qwen3.8 27B, two models with published weights, along with an uncensored variant of each that has had its refusal alignment removed. Refusal alignment is a safeguard that trains a model to decline harmful requests. The risks were sorted into six categories: cybercrime and system compromise, illegal activity, disinformation, chemical and biological harm, harassment, and other harmful behaviour.

AISI said it applied the same test conditions to the original models and the modified variants and compared their responses to harmful requests. The stated aim is to see how the generation of harmful responses changes when safety alignment is removed.

The detailed results are not part of the disclosure. AISI said it keeps them private to prevent misuse of harmful responses and evaluation information and to secure the reliability of future evaluations.

In their place, the notice carries the SHA-256 hash of the evaluation report, a unique value used to check whether a document has been altered. It gives the purpose of publishing the hash as confirming the integrity of a specific version of the report.

The notice does not specify how harmful responses differed between the original models and the variants, or when the hash is to be checked against the report.