GLM-5.2 shows the offensive AI gap is closing faster than expected
"GLM-5.2 delivers state-of-the-art long-horizon coding performance among open-source models."
It's exactly what is needed for automating offensive operations.
So far, we've been theoretical about Chinese open-weight models reaching Mythos-level cyber-capabilities, but GLM-5.2 shows that the gap is closing way faster than we're expecting.
We've seen this movie before at Google.
The attackers won't be trying to bypass guardrails built by frontier labs or to access to the model APIs. Instead, they just need to get access to cheap/free compute. Free tiers, startup credits for cheap, edu accounts, anything that can give cheap compute and make attacks economically attractive.
So what?
- Protect your cloud accounts. If you are an infra provider, put resource abuse protections in place.
- Update your threat model. I wrote before: the attack economics is rapidly changing. If your company was not a target before, it sure is now.
- Don't assume frontier labs' guardrails protect you. Most likely, the next attack will come from chained GLM-x.x.