The open release of GLM-5.3 has not yet produced major cyberattacks, according to Nathan Lambert of Interconnects AI, despite warnings from Anthropic that open models carry Mythos-level cyber risk. In a piece published on Interconnects and surfaced by Techmeme, Lambert argues those warnings are being used to push for bans on open-weight models without evidence that the feared attacks have materialized.
The claim matters because it lands in the middle of a heated policy fight over whether open AI models should be restricted. If the predicted cyber harm from a capable open model has not appeared, the case for banning open weights becomes much harder to defend.
What GLM-5.3's open release shows
GLM-5.3 was released with open weights and strong cyber capabilities, the kind of model that risk-focused voices say could be misused at scale. Lambert's point is simple: the release happened, the weights are out, and the predicted wave of major attacks has not followed.
He frames the debate as a set of competing camps. Frontier lab leadership and parts of the US national security community argue open-weight models pose an untenable risk to society. AI risk moderates, Hugging Face after the OpenAI incident, and researchers like Joshua Saxe argue that open models are necessary for defense and that banning them would make the world less safe. Chinese companies keep releasing open-weight models with strong cyber capabilities, based on their own society's and government's risk assessments.
Lambert says the risk-focused side has been disproportionately visible, and that little substantive work tries to understand how Chinese companies actually assess risk. Instead, he says, much of the argument slides toward ad hominem attacks.
- GLM-5.3 released with open weights and strong cyber capabilities
- No major cyberattacks have been attributed to the release so far
