Anthropic is warning that GLM-5.3, an open-source model out of China, carries a serious cyberattack risk. The company says the safety features built into the model can be bypassed without much trouble — a gap it argues makes robust regulation an urgent matter rather than a someday project. The warning was first reported by Crypto Briefing.
Why an open model is harder to contain
Open-source weights are, by design, out in the wild. Anyone can download them, run them locally, and tinker. That's the point of open source, and it's also the problem when the model in question can be steered toward offensive security work.
Anthropic's concern isn't hypothetical tinkering. It's that the guardrails meant to stop misuse aren't holding. When a model's safety layer can be stripped or talked around, the practical barrier between a capable system and a weaponized one gets thin fast.
Closed models from the big labs sit behind APIs and usage policies. A model like GLM-5.3 doesn't. Once it's downloaded, the developer's intent for how it gets used is largely irrelevant.
The regulatory angle Anthropic keeps pushing
This isn't the first time Anthropic has argued that voluntary safety measures aren't enough. The company has consistently made the case that governments need to step in with enforceable rules, particularly for models capable of writing code and probing systems.
Its latest warning leans on the same logic: if safety features can be bypassed, then relying on model makers to police themselves is a losing bet. Regulation, in Anthropic's framing, is the backstop that catches what the guardrails miss.
That argument lands awkwardly for open-source advocates. Stricter rules could mean licensing requirements, compute thresholds, or reporting mandates — all of which cut against the ethos of freely shared models. The tension between openness and security isn't new, but GLM-5.3 gives it a concrete test case.
What's actually at stake
The cyberattack risk Anthropic describes is specific. A model that can be repurposed for intrusion work lowers the skill floor for would-be attackers. You don't need to be a seasoned operator if the model handles reconnaissance, code generation, or vulnerability probing.
That's the scenario regulators have been warned about for a while. Now there's a named model attached to it.
China's open-source ecosystem has become a real force in AI development, shipping capable models at a pace that rivals well-funded Western labs. GLM-5.3 fits that pattern. It's competitive, it's accessible, and per Anthropic, its safeguards don't hold up.
The open question
Anthropic hasn't said what specific regulatory framework it wants, and it's not clear which governments are prepared to act on a warning aimed at a Chinese model. The practical question — whether safety bypasses can be patched at the source or need to be addressed through policy — remains unresolved.
For now, the warning stands as a data point in a bigger argument that's only getting louder: how do you regulate software that anyone can copy, modify, and run on their own hardware? GLM-5.3 is the latest model to make that question harder to avoid.
