pradeep1177

Big labs don’t expose logits, obviously. So I was wondering, if you wanted to play with distillation at scale, what other signals could you get from these models? Ha fun building this experiment: give Claude Code/Codex an enforced “workbook” and ask them to write down decisions, alternatives they considered, tradeoffs, etc., while working Just Normal model output, not hidden reasoning. Interesting to think about whether traces like these could be useful as a signal.

There is a big difference IMO. transcript shows a summary. The workbook shows you generated text, verbatim.

These logs containing opaque blobs could accidentally contain secrets, the researchers decoded many of reasoning blocks from public repositories and reported finding PII and credentials.

I was experimenting a bit how I could block these using an ingress path. GitHub /softcane/hamza

Interesting take. I also looked into this issue from a different angle https://github.com/softcane/hamza

And what are those necessary controls?

These data retention contracts are black boxes; you never know how your IP was leaked, and it could end up in the model's training data. It's like trusting META with your privacy settings.

I opened this feature request https://github.com/anthropics/claude-code/issues/68284 and hacked around a solution that I thought wasn't so hard to bake into the Claude code.

There are other similar open issues.

However, my thoughts are that I think this fix as a native feature in Claude code is not in the best interest of Anthropic's income statement. For example, many regular users are currently not utilising the weekly quota usage and that unused capacity is shared among others who are either enterprise users or know how to precisely set up the loops.

Imagine this, if this becomes a native feature, developers would start scheduling tasks at 9 pm and would continue until they wake up.

What you think?

If it’s a preference question then certainly yes, but how to push back on x productivity boost argument when a company enforce this as policy?

Yes a soft policy but it’s a smart move to encourage the rat race.

Author here, I would happy to answer your question's

Then why to even keep codex open source?

Many times those transcript contains sensitive credentials, so it’s redacted?

Im working on a project to prevent Claude code from reading any secrets key

But the whole point of this is to prevent the distillation and identify the list of blocked providers. If a provider is capturing the proxy, they can identify and modify that as well, so it only looks legitimate to the model. What am I missing here?

I used my proxy https://github.com/softcane/cc-blackbox setup to capture this.

This is how it looks.

# userEmail The user's email address is <my email>. # currentDate Today's date is 2026-06-30.

      IMPORTANT: this context may or may not be relevant to your tasks. You should not respond to this context unless it is highly relevant to your task.

</system-reminder>

I also do not understand what's the point of this, because if I have a gateway that can detect it, then we can replace the text before forwarding to the model, so what's the catch?

I generally believe the proxy route is best to understand any harness. I been building some thing similar.

So, how are you handling read/write caching? I mean, if I keep routing the next prompt based on the task weights? How about if I'm sending every 5th query to opus, which do expensive write cache?

I've been pondering over the same question, but also on an enterprise scale. For example, how could I scrub the customer data, like name, domain, etc., before sending it to LLM if my Claude code is using some MCP to fetch and debug an issue?

I mean, hooks are nice, but they are messy to manage/operate for many developers at a time.

https://github.com/softcane/cc-blackbox I started investigating the proxy path, which I think would be useful for such scenarios.

I think the expectation that AI will be the silver bullet for the domain where human behaviour change is just futile.