My approach to this has been a NixOS host with the harness running in a rootless podman sidecar.
The host has squid configured with a self-signed CA and networking rules to route all host traffic to the intercepting proxy, so I have a tight firewall and full auditability.
Then there’s a python rpc daemon running on the host with a set of whitelisted commands, read-only for pulling logs and diagnostics.
By default, the agent runs in a split pane tmux session with a host shell on the left and the chat interface on the right. The rpc whitelist includes the proper `tmux capture-pane` invocation to pull from the host shell, so I can easily let it see what I’m doing if I want it to help debug something.
I’m using pi as my harness and have custom extensions that give Yes/No confirmation gates for any writes the agent makes and that pass all bash commands/file writes to a deepseek subagent for review.
Still early days, but as someone with a similarly paranoid mindset around running LLMs securely, I think the future is promising and we’ll see some new “best practices” and related tooling popping up shortly.
Unfortunately, they’ve been suspiciously quiet since their initial announcement, but fingers crossed. Might actually have a viable new car option if they’re successful.
(Virtual card provider that generates cards as a free-to-the-user service. They make their money from a cut of the standard transaction fees. Cards are locked to a single merchant and it’s easy to configure limits.)
This is a horrendously bad-faith take. You know full well it’s *not* just a one-off $200 issue: they treat customers like this at scale.
Don’t pretend this is an isolated matter, or that CS/billing is the only arena where Anthropic has such systemic issues.
I don’t know you, but your response honestly reads like it’s coming from someone wrestling with their own moral compromises. If so, please take a good hard look in the mirror. (E: yep — https://news.ycombinator.com/item?id=47953576)
What people typically mean by the GP statement is that the “thinking” mode of these models is loosely analogous to what humans do: a bit of a retrograde reconstruction of how we arrived at a gestalt conclusion that sounds good, but may not accurately reflect the real logic at play.
IME you can see this more easily with less-polished models like Deepseek 3.X, where the reasoning in the thinking traces occasionally contradicts or has zero bearing on the non-thinking output.
> They were shirtless, unarmed and carried no visible communications equipment. They also appeared to have no idea what had just hit them, or that the U.S. military was weighing whether to finish them off.
> "The video follows them for about an hour as they tried to flip the boat back over. They couldn't do it," one source said.
Cmon now — I’m a rabid privacy nut but this is unfair given the context of:
> Prompt telemetry is opt-in and off by default. The hook asks once; if you don't answer, session-end cleanup marks it as disabled. We don't collect prompt text unless you explicitly say yes.
The UUID part is just one accessory layer, and something plenty of other players in the ecosystem don’t bother to stick to.
Feels like actually bothering to ask users for consent is what got them burned here, when I’d say it’s at least an improvement that they’re asking at all. Many products don’t, and users never bother to turn it off because they don’t know and don’t care.
I think this whole UX is deeply misguided but at least has plausibly benevolent intent.
> I'm sorry YT isn't working well for y'all. That said, expecting some grunt employee like me to feel personally responsible (or even "ashamed"??) is ridiculous.
The point of a company is what it does: The money in your paycheck comes from anti-competitive behavior, denying accountability and customer support to your users, and yes, even the enshittification of Youtube.
Deny it all you want, but you have every opportunity to walk away and do something better for society with your life if you so choose.
Internet or not, I post under my real name on here, and I fully stand by my words. Anything I say on here, I’m 100% willing to say to someone’s face. We can link up for coffee or a beer next time I’m in CA if you’d like, and I’ll prove it.
The root comment is an aggressive affront to the audience’s collective intelligence. You’re in full “rules for thee; not for me” territory, and undermining your own site guidelines if you wanna let the root comment stand unchecked but go after the rightful callouts, in my book.
At the end of the day, if you want intellectual curiosity and openness, bad-faith dishonesty needs to be weeded out; thought-provoking and honest conversation should be promoted, regardless of where the contributor is employed.
The problem isn’t working for Microsoft. The problem is dishonesty.
You’re treating the root comment with kid gloves because it’s from a Microsoft employee. Please don’t do that.
> The impulse to hit back against what is perceived as a "transparently dishonest corporate actor" is natural and human.
Honest question: If we agree that the transparent dishonesty and the lynch mob behavior are both undesirable, how do you think the two should be balanced in operative terms?
I don’t want to put words in your mouth — but are you saying you won’t allow direct pushback to dishonest corporate actors??
My view is that healthy discourse requires balance and proportionality: flagrant dishonesty, as is the case here, should license a proportional degree of pushback.
I don’t agree at all that “nobody believes this” is quite the personal attack you’re making it out to be, but I don’t care to debate that at length either.
GP did not personally attack or denigrate the person they were replying to.
As the dozens of other comments show, the overwhelming majority of us do not believe the root commentors claims, and this PM quite objectively does not have the leverage and authority to back their claim that they won’t let this happen again.
It’s hard not to read your conception of “trying for something different” as granting undue credulity to a transparently dishonest corporate actor.
Sure looks to me like this whole case is Claude Code chasing its own tail, failing to debug, and offering to instead generate a bug report for the user when it can't figure out a better way forward.
Maybe even submitting the bug report "agentically" without user input, if it's running on host without guardrails (pure speculation).