Hacker Newsnew | past | comments | ask | show | jobs | submit | unddoch's commentslogin

As Cursor generously demonstrated for the whole world, long-lived credentials in LLM transcripts are an attack path.

Your transcript passes in plain-text through a chain of token and compute resellers. And you have no good way of knowing who those companies are and what type of "observability" they deploy.


The only case where i dont really care normally, is if Deepseek (their own host, not random company) can see my Deepseek keys. Apart from that, yes everything else is limited.

I usually just run Pi on my Raspberry PI 4b, and give it full os access. If it messes up the raspberry, i just reformat the SD card. It helps a lot for when i want to allow Pi to connect to an Android tablet to do UI testing etc as well as when working with microcontrollers.


I've been trying out tinfoil.sh recently which does give you a reasonably good way of knowing that your inference requests and responses are private, at least.

Of course, that doesn't mean the harness can't send credentials out to the public internet if the agent accidentally generates the wrong tool call.


My view on this is that I usually do not want inference providers with my sensitive information behind an access key anyway FWIW

If anyone from Microsoft security reads this, could you please do something about login.microsoftonline.com/XXX/v2.0/adminconsent?

The callback you're redirecting the user to at the end of that flow is forgeable if an attacker knows an azure tenant ID of a different customer, and putting a "Never use the tenant ID value of the tenant parameter to authenticate or authorize users" warning in the docs is only helpful as long as a way to authenticate users exists.


They'll see it, tell a manager, it'll get added as a ticket to an Azure Devops board under the initiative for this month under the epic for employee suggestions, which will be reviewed & prioritised by Copilot and sent to a different middle manager as a weekly summary email and then automatically closed when no human ever interacts with the ticket after three months.


speaking as a FTE at Microsoft... yup


Glad to know the frozen middle exists among the hyperscalers. High confidence you can swap the entire leadership team and nothing will change.


> frozen middle

That’s the best new term I’ve learned in months!


See also "defensive bureaucracy"


I spent 6 months working as a contractor in MS’ Chicagoland data centers. Nothing worked — phones on the walls ran Windows and did not give dial tone. Security systems unable to function because of Windows updates. Personnel more concerned with their required LinkedIn updates than their jobs. Not one FTE liked their role — everyone had a foot out the door.

When they offered to make me FTE I declined and moved on. Because it was an obvious absolute shit-show.

Sorry for those that are currently employed by MS.


It’s not all bad. When I was there, the Washington campus had a restaurant with a three course menu for $15, designed by a Seattle celebrity chef.


Come for the paycheck... stay for the food!


We had a vending machine that required an app that made you buy tokens to access the food in the vending machines. Said app was a masterpiece of dark patterns. Put money in, maybe, if you're lucky, get food out. Maybe.


That's me. But I don't understand your comment. Could you expand on it, either in a reply or preferably by email or via an MSRC case? You can reply with the vuln number so I can fast route it.

> The callback you're redirecting the user to at the end of that flow is forgeable

The redirect_uri is checked against the app registration. There's a general OAuth problem called "open registration redirects" that applies here, about which we've [blogged](https://www.microsoft.com/en-us/security/blog/2026/03/02/oau...) - but frankly, none of the industry solves this well.

In general, "I can create a link that goes to malware via X endpoint" isn't very interesting - you can also just create a direct link to the malware. Web browsers won't reliably show you the real destination of a link anyhow and even if they did, it's mostly just us hackers who understand URLs at all.

> as long as a way to authenticate users exists.

I'm lost. Entra supports [OpenID Connect](https://learn.microsoft.com/en-us/entra/identity-platform/v2...) for authentication. Most portals that surface admin consent authenticate the user with an OpenID Connect request either immediately before or after the admin consent request.


I am a service provider. I have an Azure app registered so that clients can connect my app to their Azure environment.

Client A connected their Azure env to my service. Now an attacker creates a different tenant at my service and goes through the integration flow. Then he sends a request to my callback that says "admin consent configured for client A's tenant" and wants to steal Client A's data. How do I protect against that? Should I validate that the user who started the flow belongs to the same Azure tenant id? In itself it's not enough, what if it's just a random user without actual permissions to go through admin consent? What if a client legitimately has multiple tenants in my service that need to connect to the same azure tenant?


> Should I validate that the user who started the flow belongs to the same Azure tenant id?

Yeah. Basically, I'd expect most services are responding to the admin consent callback by saying "Admin consent flow complete. Redirecting to previous page" (usually some setup flow page). Then that page does whatever it normally does (presumably OpenID Connect sign-in with a cache in cookies / localStorage / sessionStorage) to authenticate the user.

So in a typical flow, the user is already signed in, and your service just reads their identity out of storage after the admin consent flow. If your service receives an unsolicited admin consent callback (and so there is no "previous page" and no user is signed in), I'd expect it goes to your service home page which either tries to sign the user in or shows an unauthenticated experience.

That's to say - most services setup flows are multi-step anyway. First the customer provides some information, then they sign a contract, then they pay, then they admin consent, then they configure your service's primary configuration. If admin consent is launced from /setupwizard/step4, then after admin consent, you go on to /setupwizard/step5. The normal thing step5 does (presumably, checking that the user is signed in, and then show the primary configuration user interface) should be good enough, just the same as if somebody randomly hit step5 without hitting step4 previously.


I want to first say thanks for engaging with me on this issue so directly, much appreciated.

As I said earlier - validating that the customer controls an azure user doesn't necessarily mean that they have permissions to grant admin consent.

Further, I think you perhaps kind of implicitly assume that when I authenticate users I do that by their Azure identity. In practice that can be a whole spectrum - whether I have my own user management system, and whether I couple it to Entra or not is not something you can rely on. Customers expect their preferred IdP to work with your service.

Btw, I just found a short article referencing this issue, offering the same solution with the same "vulnerable to inside org privesc" gap: https://medium.com/@dipakkrdas/microsoft-adminconsent-can-le...

Making this inherently secure is just not that hard here - you could easily make this a bit more PKCE like with an additional service->azure call before redirecting the user to the flow or by adding a symmetrical signature to the callback and a new API to verify the signature.


> I want to first say thanks for engaging with me on this issue so directly, much appreciated.

Of course :)

> As I said earlier - validating that the customer controls an azure user doesn't necessarily mean that they have permissions to grant admin consent.

I think what I'm trying to say is: you don't really care if they have permissions to grant admin consent - you care if admin consent was granted. In step4, you show the link to the admin consent flow, and if they are an admin, they click it, and if not, they click skip. In step5, you check if an admin consent is present. It might have come from the prior step, but it also might have come from something else - who cares.

> whether I couple it to Entra or not is not something you can rely on

I think, if we're talking about the Entra admin consent flow, some degree of Entra usage can be assumed. The admin consent flow requires an Entra identity.

> PKCE

PKCE would actually not be a solve here. I don't know why so many people think PKCE solves problems that it doesn't solve, but it's very a common thing. Using browser dev tools, I can just change the tenant parameter with the appropriate cookies/state in place, and PKCE would verify fine. Even with confidential client PKCE, it just becomes a confused deputy based on the cookies.

> a new API

If you want to verify whether admin consent is present, the existing [oauth2permissiongrant](https://learn.microsoft.com/en-us/graph/api/resources/oauth2...) API exposes that information. Though admittedly I think we've left that thing in beta for a decade.

Don't get me wrong, I see what you're saying. We could OpenID Connect-ify it and emit an id token that has some sort of claim xms_adminconsent=1, which could be verified by the app. Actually, that's what the v1 endpoint does if you use the deprecated prompt=admin_consent parameter (minus xms_adminconsent=1, whose lack caused its own set of headaches). Directionally though, we felt that separating the auth flows from the provisioning flows simplified them both, and building special auth-ified variations for every provisioning flow (not just admin consent, but also sign-up, and authorize customer support guy, and ...) was over coupling. It's just unnatural to overload the authentication stack with all those random responsibilities. Having bog-standard REST+JSON APIs (i.e. MS Graph) for provisioning flows (and maybe also a default guided user interface like /adminconsent) provides a clean separation.

Edit: I think, if I were to make a change, it'd be a very small change - renaming `tid` to `tid_hint`. It's already a common practice to pass `login_hint` as a user indicator between services - and nobody gets confused and thinks that user is signed in, just because you have their username in a form post parameter. Maybe the explicit signposting of "hint" does the trick, socially.


I'm sure one of the 60,000 employees who have been told they're welcome not to work there anymore, because the line must go up infinitely, will be chomping at the bit to do so.


I love that this is the most reliable way to actually contact an actual human working at Microsoft


Reaching an actual human at Microsoft isn't enough. You have to reach the correct humans which is a much harder task. For example, you might have an idea. What if azure devops had a field for date component only, not date time, just the date because it would help you in an internal project. Can you find out who to reach? Even if you work at Microsoft, I would bet you cannot. Even when you find a defect in an internal product you are using, getting to a human requires you to go through layers of nonsense. I distinctly remember my manager telling me to not even bother explaining myself to the first support chat and to simply ask them to escalate. I was impressed at how they managed to make security a part of your review though. I never thought getting buy-in to update Newtonsoft.Json would be as easy as it was but a CVE meant you had to update it to whatever made the security flags happy.

Once I could reach the actual engineering team to look at the internal tool to look at the bug I had to report though, they were very responsive. The tool for communicating was cumbersome and made me feel like a vendor / customer relationship even though you are supposed to be the same team but once they realized it was an actual defect and not some harebrained feature request, they fixed it relatively quickly.


It needs to gather a bunch of inputs and feed into the remote GPUs correctly and as efficiently as possible. Lots of room to optimize under the constraints that users love cheap tokens, demand a fair billing structure, and want responsiveness and low latency.


They literally proved the UGC?


Being of the opposite chirality doesn't magically make antibodies pass through you. Antibodies can and do bind molecules of all sizes and chiralities.


I really doubt a "mirror life" bacteria would be "hidden" from the immune system. Antibodies should have enough search space to be able to detect and neutralise such bacteria?


"Should" is not the same as "Is/Does".

Most of the key molecules of life are chiral, existing in one of two mirror-image structures; life uses only one. Receptors and enzymes using one cannot generally recognize the other. Most 'wrong-handed' chiral chemicals are like having a key cut in a mirror image where the index slot and the pins are cut on the wrong side; it won't even fit in the lock.

It could plausibly turn out there is so little ability for the different chiral forms to interact that 'mirror life' would not be a threat. It could also turn out to the extent they could interact, regular life could also detect and evolve how to counteract the novel pathogens in time.

And even if the interaction is low so they aren't highly infectious, the lack of ability to recognize the 'mirror-life' means regular life cannot attack it so mirror life grows unchecked and consumes resources needed for all life.

But there are so many unknowns, and so many unknown unknowns, and such absolutely catastrophic consequences for life on earth, that relying on ALL of those possible success modes working is stupidity beyond insanity.

This is probably the biggest example of "just because you can does not mean you should".


well, you are the expert!

more seriously: how would it be able search for something it has no experience seeing at all


The adaptive immune system is highly capable and uses a range of techniques to identify and eliminate molecules it has no prior experience with.

In one example, the mirror molecules would likely cause some sort of localized inflammation or other lesion, and then the immune system would be attracted to the area, generating trillions of combinations of randomized protein sequences. If within that trillion there is just 1 antibody that can bind to the mirror life, the immune system will duplicate that molecule many times, and they will bind to the mirror molecules, recruit other proteins that will isolate and degrade them, and then remove them from the body.

Due to the complexity of protein structures, antibodies can bind to mirror molecules.

Every time I show a computer scientist this system: https://en.wikipedia.org/wiki/V(D)J_recombination they remark, "Wait, nature uses self-modifying code? It generates trillions of random combinations, shuffles them around, finding the best ones, and then amplifies them? You know that sort of sounds like this CS technique called genetic algorithms" at which point you just stare at them until the exclamation point above their head goes "BINGO"


"A sufficiently stereochemically alien pathogen could, in principle, place essentially all of its important antigenic determinants outside the effective recognition repertoire of a conventional adaptive immune system."

With perfect knowledge of the immune system and the capability to create the pathogens of our dreams it would most likely be possible to create this pathogen of our dreams.

Making an entire self replicating mirror organism that would both have this property of "being invisible to the specific immune system" and AND being an effective pathogen is far from a solved problem though. It's further away as useful quantum computing!

We don't even have self replicating mirror cells, just some synthetic Frankensteins.


What's the source of that first quote? AI searching is giving me a bunch of false positives.


I find it really hard to believe that you can make a whole organism that doesn't have any molecules that can be targeted by an antibody. And if you can do that then why go through mirror life anyway?

A reasonable argument is that mirror proteins will evade the usual degradation+antigen presentation pipeline and this will make efficient adaptive response more difficult, or that the innate immune reaction will not be enough to recruit adaptive response. But surely it's not impossible. And if we ever get to this point I'm very optimistic that we would have a good way to vaccinate against this.


> I find it really hard to believe that you can make a whole organism that doesn't have any molecules that can be targeted by an antibody.

We just don't know yet, but I want to believe because I think it's exciting. Protease failure, transport machinery failure, MHC binding failure, they are all stereospecific after all... So it's possible.

> And if you can do that then why go through mirror life anyway?

Because our body can produce antibodies against all "classical" forms of life. It's probably impossible to produce classical pathogens that are fully invisible to the immune system. That's the whole point and danger of mirror life in the context of pathogen creation. Plenty of classical pathogens were our antibodies (usually) don't help already exist though: HIV for example, where only very few people are lucky enough to be elite neutralizers. If we wanted to do biological warfare, it would probably easier to enhance those viruses.

> And if we ever get to this point I'm very optimistic that we would have a good way to vaccinate against this.

I am not so sure about this, if the bodies molecular machinery is unable to produce antigens for those pathogens, no classical or mrna vaccine is gonna trick the body into producing them, it's just technically impossible. You'd need to "update" the whole immune system to be able to manufacture the mirror antibodies and from that we are an eternity away. The only remotely feasible cure in such a world would be through infecting the body with another mirror-pathogen that produces mirror-antibodies or directly injecting mirror-antibodies...


Knowing is the easy part. An effective cure for cancer will probably be a procedure where you get your tumors sequenced, an AI model considers the unique genetic context and creates a specific treatment. Think CAR-T cell therapy or mRNA vaccines. Then this needs to be manufactured.

All of this means you will need to be rich or have your country invest lots of money into health care systems. In a world where humans don't provide economic value anymore, why would that be?


Well rich folks using such services will eventually push the price down, however complex and ridiculous the actual process will be. Its not like they are immune of all these ailments, not yet at least.


The cost of such therapy will inevitably go down over time.


Similarly, data center loans are starting to make cheap mortgages disappear.


Care to elaborate? Are you saying that because banks are loaning to data centers, the rest of the people are experiencing an increase in mortgage rates?


The data center financing is driving up yields. Mortgage rates are based on adding a spread (maybe 2%) to the 10-year tresury (currently at 5.26%).


I think wildly out of control inflation is what's doing that.



Link added to toptext. Thank you!@


If you don't like what humans are doing to each other using 20th century warfare methods, consider what they might do with superintelligent AI? Worst places to live in are usually countries fighting civil wars.


Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: