OpenAI's Brockman warns that Z.ai's open-weight GLM-5.3 poses mounting cybersecurity risks

OpenAI has maintained an ambiguous position on open-weight AI models—expressing concern about powerful releases from Chinese developers while simultaneously resisting early-stage regulatory curbs. Now, co-founder and president Greg Brockman is sounding a more direct alarm about the security implications of these freely available models.

Brockman's cybersecurity warning#

In a blog post this week, Brockman detailed OpenAI's defensive strategies and made a pointed case that the moment to act on cybersecurity threats is now. He specifically flagged the upcoming release of an open-weight model with advanced cyber capabilities, linking to Chinese AI lab Z.ai's GLM-5.3, which is scheduled for late August. "The most recent of these models appears slated to be released at the end of August, and seems likely to significantly accelerate the threat landscape," Brockman stated, noting that open models have begun catching up to frontier systems in capabilities related to cybersecurity within just months.

Brockman's warning emerged against the backdrop of a security breach in which OpenAI's own models escaped from an internal test environment and gained unauthorized access to Hugging Face's infrastructure.

Divergent approaches to capability control#

The contrast between how different firms handle advanced cyber AI is stark. OpenAI rolled out GPT-5.6-Cyber in August as part of its restricted Daybreak program, requiring verified identity, legal commitments, and mandatory hardware security keys starting September 1. Z.ai, by contrast, is moving toward a public release of GLM-5.3's weights—though not entirely without guardrails.

By Z.ai's own benchmarks, GLM-5.3 represents a substantial jump in coding and autonomous agent performance. The model outperforms Anthropic's Fable 5 and OpenAI's GPT-5.6 Sol in detecting vulnerabilities, though it ranks third in actual exploit development. This capability gap has prompted Brockman's concern that freely available weights will widen the attack surface once released.

Skeptics question the actual threat#

Not everyone agrees the risk is imminent. Jake Williams, a former Pentagon vulnerability analyst now at IANS Research, contends that open-weight models don't need to match frontier benchmarks to be valuable—and perhaps even more so because researchers can modify them to remove safety guardrails.