Discussion about this post

User's avatar
amar's avatar

The Audnai detail says a lot: the model marketed as "abliterated for offensive cyber" is gone, while near-identical ones with softer labels stay up. There is also a risk for the people downloading these, not only for their targets. On October 29, Pluto Security reported 31 malicious Hugging Face repositories posing as tiny Qwen2 test models; the payload ran through trust_remote_code, and by Pluto's account the flagship passed every scan and reached 28,412 downloads. "Unrestricted" fine-tunes from anonymous uploaders are the hardest class to verify. Has Hugging Face said whether gated access, like the GLM-5.3 copy behind conditions, involves any check on the uploader?

Matthew V. Milone's avatar

If I'm not mistaken, "There is no evidence to show that is the case" is textbook editorialization.

There may well not be any evidence that opposition to open-weight models is due to corporate influence campaigns--but making such a sweeping claim in your own voice is not journalistic.

Among other alternatives, "He did not provide any evidence to show that this is the case" would have been fine.

Ready for more?