A Chinese startup is about to release an open-source AI model that it claims already matches its American rival model at Anthropic.
Zhipu AI (Z.ai), a Chinese artificial intelligence firm with ties to the Chinese Communist Party (CCP), allegedly exceeded Anthropic’s ability to find software security flaws after Z.ai’s GLM-5.3 model scored higher than Anthropic’s Mythos 5 model in a test environment, according to Reuters. Both models are currently restricted from public access due to cybersecurity concerns.
An unverified CyberGym test of the models’ coding review capabilities found that GLM-5.3 scored 84.5 percent, while Mythos 5 scored 83.8 percent, the outlet reported. GLM-5.3 was less efficient at turning security flaws into real software exploits, scoring 54.4 percent, significantly lower than Mythos 5’s score of 78%.
Z.ai claimed its model will be publicly available in roughly two weeks following a safety assessment, but that its risky cybersecurity abilities would still remain restricted to a “trusted access” program, Reuters reported.
Anthropic launched Claude Mythos 5 to limited private partners in June, alongside a public version, Fable 5, that featured safety measures, according to the company’s June 9 announcement.
“Claude Mythos Preview, and now Claude Mythos 5, are models with significantly stronger cybersecurity capabilities, especially in exploit reasoning,” Anthropic said in a statement on Claude’s cybersecurity page. “As this capability carries the greatest potential for misuse in security, we are limiting initial access to a small number of partners through Project Glasswing.”
Z.ai and Anthropic each did not immediately respond to the Daily Caller News Foundation’s requests for comment.
“Introducing GLM-5.3: Built to Code. Ready for Cyber Defense,” Z.ai said in an X post announcing its new model.
The Chinese startup clarified that the new GLM-5.3 model will only be released to a small number of partners when the product is first introduced.
“An initial group of partners is now offering GLM-5.3-powered services through our official service, with its safeguards and usage policies in place,” Z.ai said in another X post in the announcement thread. We’re expanding partner access through a consistent and responsible process and will share updates publicly.
“Zhipu Al maintains strong links with China’s government and state-owned entities, as evidenced by over $1.4 billion in state-backed investment, numerous government procurement contracts, and active involvement in setting national AI technical standards,” OpenAI said in a June 2025 Substack post, claiming the Chinese Communist party often works with Z.ai’s leaders.
Z.ai, was founded in 2019 and originated at Tsinghua University. The company’s mission is to “[t]each machines to think like humans, benefit humanity with reliable AI,” and drive Chinese innovation, according to Z.ai’s “About” page.
Hugging Face, a U.S. AI firm, protected itself from an OpenAI agent cyberattack in July with Z.ai’s GLM-5.2 model, a July 27 Hugging Face blog post said.
(DCNF)
