Chinese AI Model GLM-5.2 Matches “Dangerous” Capabilities of Anthropic’s Claude, Experts Warn

The rapid advancement of AI models in China is drawing significant attention, with recent analyses indicating that the GLM-5.2 model from startup Zhipu.ai (Z.ai) possesses cybersecurity capabilities comparable to leading Western AI systems like Anthropic’s Claude Opus. This development is particularly noteworthy given the current US government’s restrictions on access to advanced AI models due to potential misuse concerns.

Independent cybersecurity assessments have positioned GLM-5.2 as a powerful tool for identifying system vulnerabilities, with some experts drawing parallels to the capabilities of Anthropic’s “Mythos” model, previously deemed too dangerous for public release. This comes at a time when the US has placed temporary bans on certain frontier models from Anthropic and OpenAI, creating a unique geopolitical moment in AI development.

Key facts

The AI and Cybersecurity Threat

The discourse surrounding AI’s potential as a weapon intensified with the emergence of Claude Mythos Preview. The realization that these models could be used to discover and exploit system vulnerabilities led Anthropic to restrict Mythos’s public availability. Despite the release of scaled-down versions like Fable 5, US authorities have imposed temporary restrictions, a move that has also affected GPT-5.6 models.

Chinese Cybersecurity Tools Emerge

In parallel, Chinese cybersecurity firms are also making strides. 360 Security Technology (Qihoo 360), a company listed on the US “Entity List” since May 2020, recently launched Tulongfeng, a vulnerability detection tool. According to its creators, Tulongfeng is comparable in capability to AI models like Mythos. The company’s CEO, Zhou Hongyi, has previously described advanced AI models as akin to “cyber-nuclear weapons.”

GLM-5.2: A “Mythos” at Home?

Zhipu.ai’s GLM-5.2 is gaining prominence for its strong performance, challenging the dominance of US-based AI companies. A significant advantage of GLM-5.2 is its open-weight nature, allowing for widespread download, modification, and execution on private hardware, though it requires substantial video memory due to its 744 billion parameters. Beyond programming and agentic tasks, its prowess in cybersecurity is drawing particular attention.

Semgrep, a cybersecurity firm, recently analyzed GLM-5.2 and reported that it surpassed Claude Opus 4.8 in cybersecurity tasks, with researchers noting, “We have a Mythos at home.” A separate study by Graphistry corroborated these findings, comparing GLM-5.2 favorably against Opus 4.8 and GPT-5.5. Notably, GLM-5.2 achieved these results at a fraction of the cost, reportedly one-sixth of the expense incurred for Claude Opus 4.8 tests. An Axios report cited a cybersecurity researcher who stated that GLM-5.2 can chain exploits in a manner similar to elite human attackers.

China’s AI Timeline

The pace of China’s AI development has sparked debate about future timelines. In response to speculation that China would have a Mythos-level AI model by the end of 2026, Z.ai CEO Jie Tang confidently stated that it would take less time. Elon Musk had previously predicted the arrival of such a Chinese model in the first quarter of 2027.

Sakana Fugu: An Orchestrator Model

Meanwhile, Japanese AI startup Sakana AI has released Fugu. This system is described less as a standalone AI model and more as a router or orchestrator designed to leverage various open and closed models to achieve performance comparable to top-tier US models. While internal benchmarks are promising, some independent analysts suggest that its performance and cost-effectiveness may not fully align with the company’s claims.

A Paradoxical Moment for AI

The current landscape presents a paradox: as the US restricts access to its most advanced AI models to mitigate cybersecurity risks, Chinese companies appear to be rapidly closing the gap with powerful, openly available models. This dynamic raises questions about the future of AI development, regulation, and global technological competition.

Source: “Tenemos un Mythos en casa”: los expertos ya afirman que el modelo chino GLM-5.2 es tan “peligroso” como el de Anthropic, Xataka, https://www.xataka.com/robotica-e-ia/creadores-modelo-chino-glm-5-2-afirman-que-igual-peligroso-que-mythos-eso-plantea-momento-sputnik-para-ia-china