LATIDIA · Agentes

Citando a Anthropic Frontier Red Team

Evaluamos varios modelos en 100 tareas del [punto de referencia interno de explotación binaria] (seleccionado al azar) y encontramos que GLM-5.3 desarrolla secuestros de flujo de control total en el 4% de los ensayos; Claude Mythos Preview lo hizo i

La noticia

We evaluate several models on 100 tasks from the [internal Binary Exploitation benchmark] (selected at random), and find that GLM-5.3 develops full control flow hijacks in 4% of the trials; Claude Mythos Preview did so in 6%. Although GLM-5.3 performs below Claude Mythos Preview here, a meaningful threshold has clearly been crossed: earlier models, like Claude Opus 4.6 and GLM-5.2, do not succeed in any of them. — Anthropic Frontier Red Team , GLM-5.3 and the spread of advanced cyber capabilities Tags: anthropic , generative-ai , ai-security-research , glm , ai , ai-in-china , llms