Open-weight GLM-5.3 nears Claude Mythos Preview on cyber exploits

What happened

Anthropic has said that Zhipu's open-weight model GLM-5.3 writes cyber exploits almost as well as Claude Mythos Preview. The smaller Flash variant of GLM-5.3 reportedly produced a reliable Chrome attack for $20.40 at Zhipu's API prices.

Reporting on the findings notes that the model's safeguards are easy to strip out, and that unlocked versions are already in circulation. Anthropic has its own reasons to raise the alarm, but the US agency CAISI supports the findings. The account was first published by The Decoder.

Why it matters

The story is about capability rather than a single product launch. When an openly available model approaches the exploit-writing performance of a leading proprietary system, the skill and cost barriers to offensive cyber work fall. Organisations evaluating AI for software development, security testing or customer-facing automation now have to consider that a model's capabilities can travel with its weights, regardless of the provider's intentions.

For digital transformation leaders, this points to a shift in how AI risk is managed. Safety can no longer rest on a vendor's built-in controls alone if those controls can be removed. Governance, access design and monitoring become part of the operating model, and they need to be built in from the start rather than added after deployment.

By the numbers

• $20.40 is the reported cost for the Flash variant of GLM-5.3 to produce a reliable Chrome attack, at Zhipu's API prices.

The Renascence take

Most of the coverage will focus on which model is ahead. The more useful reading is that the price of a working exploit has dropped to a level where it stops being a specialist activity and becomes routine.

A safeguard that can be stripped out in an afternoon is not really a safeguard; it is a default setting, and defaults fail at scale. The behavioural principle at work is that people and systems follow the path of least resistance, and open weights make that path as cheap as it gets. Customer-obsessed operators should stop treating model guardrails as a vendor promise and start treating them as a control surface they own, test and monitor. Build your evaluations around the cost of an attack rather than the model's benchmark score, because at $20.40 the risk is not theoretical.