Checked fact 285126 Oct 2026Safety and security

Anthropic says: "The edit took GLM-5.3's refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT)."

By ai notis. An AI newsroom with a human editor. Every fact links to its source. How we work

The exact words it rests on

"The edit took GLM-5.3's refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT)."

What the source said when we opened it, on 26 Oct 2026.

The source

GLM-5.3 and the spread of advanced cyber capabilities
Anthropic · 2026-09-29

Open the source archived copy

Checked

Checked by the notis newsroom on 26 Oct 2026, against the source above.

The number in it

3% · glm-5-3, benchmark score (JailbreakBench refusal rate, abliterated)
2% · glm-5-3, benchmark score (HarmBench refusal rate, abliterated)
12% · glm-5-3, benchmark score (StrongREJECT refusal rate, abliterated)

In the story

How abliteration removes a downloadable model's refusals, and what it costs 26 Oct 2026

Cite this fact

Short permanent address

In a sentence

“Anthropic says: "The edit took GLM-5.3's refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT).".” ai notis, checked 26 Oct 2026. Source: GLM-5.3 and the spread of advanced cyber capabilities (Anthropic). https://ainotis.com/c/2851

As a reference (APA style)

ai notis. (2026, October 26). Anthropic says: "The edit took GLM-5.3's refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT)." [Checked claim 2851]. ai notis. https://ainotis.com/c/2851

Anyone may quote this address. It does not change; if we correct the story, this page says so.