Quote:
Originally Posted by deepzero
Somehow it really doesnt feel good to depend my reverse engineering results on leaked/stolen webaccounts or relying on tricking AI censorship. Neither are reliable and can disappear overnight.
|
I was inspired by
@Shub-Nigurrath tut last year on hunting for publicly exposed Ollama LLM instances:
Quote:
|
https://forum.exetools.com/showpost.php?p=133352&postcount=17
|
From that point onward, my efforts shifted to locating the Claude and OpenAI accounts that IT‑security staff share(from companies they are employed like
KrebsOnSecurity etc).
I want to emphasize that we never hack any military servers or websites. We simply use the frontier‑AI accounts that the military and other government agencies provide, which lower‑level personnel hand out for a modest fee.
These accounts are accessed through their own proxy servers, so we have no direct interaction with the military or any government agency.
All requests to the AI providers appear to come from the reverse‑proxy IP addresses, not ours. As long as we avoid supplying any personally identifying information to the models, our identities remain protected.
As @Shub-Nigurrath noted in the post last year about hunting for unsecured Ollama servers, this practice is no more illegal than using cracked software.
Quote:
Originally Posted by deepzero
S
I think we should focus on a) less-restricted chinese models, preferable opensource ones b) smaller local models or c) running opensource models on rented GPUs online.
e.g. Exodia is running 2x DGX Spark (~12k usd): https://x.com/mrexodia/status/2090806161813405917
|
I like this approach. You can see me repeatedly ask @Shub-Nigurrath last week how we can hook up the mcp to local LLM like the Qwen models.