Using a throwaway account for obvious reasons, but I’m very involved in this space using LLMs from multiple providers. I’m aware of at least two instances in which the intermediate infrastructure “swapped” responses, once impacting Claude models and once impacting GPT models, from two different providers.
One gave us a proper postmortem in which their API gateway was incorrectly handling HTTP 100 status codes, putting them into an error state where there was effectively an off by one error - you would receive the response to the prompt that came in before yours and would pay it forward (your response would go to the next caller).
The other instance never had root cause explained to us, and we were just told to trust it wouldn’t happen again.
Both of these are from $1T+ companies.
ZDR wasn’t compromised in these cases since it was responses being swapped in flight. I wouldn’t be surprised if this is a similar issue - it’s not that data is being retained, it’s just not being safely isolated in intermediate infrastructure.
This attack is called "HTTP desync" or "request smuggling". It's often done intentionally by a client to try and spy on other clients' responses.
Every time you multiplex requests from multiple clients onto one upstream connection, you are probably vulnerable to this, because (despite its superficial simplicity) HTTP is just too complex to reliably match the requests and responses to upstream.
For example a desync can be triggered in some systems by having more than one Content-Length header, by mixing Content-Length with chunked encoding, or by passing an HTTP/2 header called Content-Length that doesn't match the actual content length.
The same attack has been applied to SMTP by messing up the line endings surrounding the end-of-message delimiter, where it's called SMTP smuggling. It may also apply to other protocols.
Very true, this was likely an attack. Worth noting that mr kettle has done a defcon talk nearly every year on some variant of this attack, the most recent one titled "HTTP/1.1 must die" because he rightfully believes that switching to the binary headers of http/2 (specifically in reverse proxy connections to upstream servers) is the only way to systematically prevent these.
Actually, it’s not obvious why you’re using a throwaway account…
Every emergent behavior from these actors - whose claim to positive moral values is barely plausible - should be reported, discussed, dissected and critiqued early and often.
Woah. Sounds plausible. However, wouldn’t that still be an implicit violation of ZDR since now the response is possibly egressed out of the enterprise network? So if I were working with PHI, the response egress is a potential violation of HIPAA even though claude didn’t retain anything — but the whole
Point was to comply with HIPAA. Thoughts?
These companies(at least one of them) seem lead by idiots(Hint:his name is Dario) so I wouldn’t be surprised to have multiple wtf moment if you were to see how they treat our data…Let’s just start pushing for opening up AI models because they are too dangerous behind paid walls. That would be a great regulation.
Dario quit OpenAI to hype the AI apocalypse for quick cash and attention. Then, he walked right into an obvious crisis with the Pentagon by continuing to try to play both sides of the AGI doom story that even his own AI would've pointed out. Then, after being labelled a supply chain risk, he starts a new roadshow with the newest most dangerous AI model that definitely cannot be released to the public and its safer little brother Fable. A move that gets both his premier models shut down globally once the same government that labelled them a supply chain risk learns that Fable isn't actually safe from jailbreaks. Just prior to his planned IPO.
Dario might not be a literal idiot, but he might strongly benefit from training a model to do strategic thinking for Anthropic.
All of these things have people frothing at the mouth to give up all their data to Anthropic to use their models and to buy in when the IPO eventually happens.
Seems to me Dario is actually a genius. These are all things that I would to make people believe that my “basically the same as the other guy” product is ackshually best thing ever for real. Trust me bro.
The entire bubble is hype and fear mongering. The technical merits of the products are completely irrelevant at this point. Dario is doing exactly what someone that understands this would do and they are winning.
HN thinks the safety crowd is dumb, and has never seriously engaged with the AI safety space.
HN doesn't believe superintelligence will be a thing; while the AI safety crowd believes they are building it. So the decisionmaking of the safety crowd is incomprehensible to HN.
Funny how Dario’s and Sam’s concern for our safety dovetails so nicely with their companies’ strategies.
How fortunate.
Grow up. Whenever push comes to shove, they reduce safety and alignment departments, rush out releases over the heads of the same departments. If you engaged with the news these last years you’d see it for what it is “models for me, but not for thee”.
It's clear you haven't engaged with the subject matter beyond the typical "internet-forum cynic" mindset.
Both companies were founded on the basis of AI Safety.
- There are tons of great safety people doing real work at OpenAI. Releases are held back, models are evaluated, etc.
- Anthropic goes even further - constrained themselves with a PBC/LTBT structure, treat safety even more rigorously, and notably delayed the release of Mythos (literally the opposite of what you alleged) and continue to hold their two red lines despite threats from the gov.
You should actually talk to some of the people at these labs. Nearly everyone working at these places genuinely believe AGI/ASI is actually happening, so they do take safety seriously.
To imply these companies don't care about safety is typical internet-brand nihilism/cynicism that helps you feel smart while being literally the opposite of the truth.
There is no reason for you to make personal attacks like that. Not on HN.
Moreover, your take on Dario is over simplistic, and undersells the extent to which Anthropic takes seriously safety. It's not lip service, there are real dollars and attention spent on alignment at Anthropic.
Thanks for the detailed report. We’re confident this is a hallucination but of course take these reports seriously and the team is looking into it. We’ll report back if anything turns up.
I know it's the weekend, so thanks for working hard. Just a suggestion from a user: I wish we could manage Claude Code's memory more easily. Right now, when I go into the .claude folder and change a project folder name or something, sometimes it can't pull up the memory properly. It'd be nice if there were an easier way to import or export it. Thanks!
I’ve been seeing this in Gemini in the past few days. Often during a prompt with a reasonably large input set, I’ll get answers that appear to belong to someone else. It may be trigger hallucination, but it seems like it may be cache collisions or something else. I’ve not seen anything to suggest private information is leaking, but it’s disconcerting to be researching something and then get what appears to be a math tutoring response.
I’ve also had problems with Gemini when accessed through their UI in the past few weeks. That’s concerning that you are also seeing it several days later in a different context.
I wonder if there could be a large security situation playing out behind the scenes right now.
I’ve been working on using AI to assist me in writing meta parsing grammars. Fortunately I have not launched most of them yet. I know for a fact that the next generation of models represent a major step change in basic vulnerability identification and exploitation, especially if you know where to point them. They’ve found several bugs and at least one exploit in my parsing tools so far, I can’t imagine how many there still are waiting to be discovered across the entire modern tech ecosystem.
My whole company is doing mid year reviews and Gemini is the only allowed tool and its been flumoxing people with seemingly random unrelated responses. Often in different languages.
That is when it bothers to respond instead of just sending back an 1099 error code
Sounds like a hallucination unless proven otherwise, even the leading LLMs can do those from time to time, and they will always appear plausible like that. Also could be the session having a lot previous context, like 800K+, which (I think) makes hallucinations more likely.
Relevant comment from the OP which makes a hallucination more likely:
> There is one tool call result that includes a string that printed a pathname including minecraft.py because it was listing the files in a Python virtual environment and the Pygments package has a lexer called minecraft.py
I realize hallucination has no precise definition but this doesn’t sound at all like anything I’ve ever heard called hallucination. Hallucination is usually plausible wrong answers or made up info that ends up fitting the most likely response (like a manufactured citation) and comes from the way LLMs work at predicting tokens. This example demonstrates completely implausible output, it’s not something that fits with hallucination.
All that said, it doesn’t require cross session leakage, it could just be training data or like those nightingale (probably the wrong bird*) data generations where they just prompt an LLM with nothing and it starts spitting out conversations.
I see a bunch of downstream comments about caching, sounds like maybe there’s an error where it loads nothing instead of the cache and so starts spitting out random generations.
* edit: it’s magpie. Worth looking at the concept, I’m not sure people realize they LLMs generate random conversations when prompted with nothing, this seems at least as likely as sessions leaking: https://github.com/magpie-align/magpie
The word “hallucination” has become overloaded, but it general means an LLM producing some output that isn’t plausible or grounded. When you have a very long context session where the context includes “minecraft.py” it’s not hard to extrapolate that Minecraft may have ended up in one of the reasoning traces and that distraction snowballed until it appeared in the output.
These effects are becoming more rare as the SOTA models are improving so much. If you spent a lot of time with earlier LLMs or you experiment with smaller, quantized local LLM models this type of thing happens very frequently. When you see it happen so much on a model you’re running on your own hardware it becomes a reflex to chuckle and reset the session with a clean context. When it happens from a hosted provider it can be scarier because it’s not the type of failure mode most people are used to seeing.
The person posting this claims to have reproduced in a separate context down the thread:
> Same thing just happened on a Claude Mobile session in same Enterprise account. Common theme in both is Sonnet 5, first response after more than 5 minutes (cache miss).
If you've never had an LLM (all models) suddenly start spouting nonsense in a completely different language...you haven't been using LLMs that much. They will go absolutely insane some % of the time.
They can “go insane” but it seems often to be infra related as opposed to anything one would consider hallucination. Smaller models will often get stuck repeating a word or phrase forever but that’s a bit different and nobody would call it hallucination.
One annoying one is we have an LLM-as-a-judge that is supposed to quote parts of a transcript to justify its reasoning, and sometimes it’ll get stuck on something short like “No.” and then just endlessly repeat it: “4. No. 5. No. […] 728. No. […] 1435. No. …”
I am facing a billing/subscription problem and there's nothing I can do or get help on. Their chatbot support shuts me down. Their email is also handled by the chatbot (not even sure whether it's the "same chatbot"). It has been a dead-end. I contacted my bank (credit card issuer) and finally a staffed said I am better off just marking the card lost and having it reissued and that's what I did in the end. I hope that works.
I've never understood in what world this world decided it was okay to hand over these much unchecked power to such corporations. But this is how it has always been one way or the other.
The person had “minecraft.py” in their context and the session context was very long.
Having an LLM session with very long context occasionally go off on a tangent is not uncommon. The people who expect absolute perfection out of every LLM interaction see this as some total indictment of the entire technology, but the people who use these tools daily have learned to treat the output as partially stochastic and to avoid extremely long context, even if the model offers it. It’s best to compact strategically or summarize next steps to hand off to a new session. Using sub-sessions can also reduce context pollution at the cost of additional token expenditure to summarize and transfer data to and from the sub-session.
Reminds me of a session I had recently (on web!) where claude insisted that i prefixed all my messages with statements about code execution or something, which was not the case. I interrogated it about that and it confirmed that it came from somewhere else, but could not get rid of it and each response mentioned that its gonna ignore those instructions. Eerie.
Anthropic injects text into the conversation triggered by certain conversation topics. This happened to me in relation to some red-teaming related discussion that was adjacent to something “sensitive”, I think sex, and Claude got confused about why I had said some kind of warning and mentioned it it’s response. After a back and forth it was clear that some extra warning to answer but avoid anything inappropriate had been inserted into the conversation.
Claude also sometimes mentions getting messages from classifiers, probably related to auto mode. Amusingly enough, when this happens to a subagent/fork, the orchestrator will call these " hallucinations by the subagent"
Not sure the relevance of this comment, but normally if someone built a classifier that bad they’d be fired. Anthropic obviously thinks they have some monopoly power they can use to foist garbage on consumers, I think they don’t.
The consequence of a too strict classifier are annoyed customers who will spend less on Fable. The consequence of a too lax classifier are export restrictions that prevent a huge chunk of their customers from using Fable
I'm annoyed but not surprised at the overeager classification
If people are complaining about Anthropic (on an only-vaguely related thread) rather than simply switching to a suitable competitor, then Anthropic clearly has some 'monopoly' power over the specific capabilities the complainer wants from them.
Fable/Opus 4.8 outperform Codex 5.5 for me at the general architecture/refactoring/performance work I’m doing, to the point where it’s not worth using Codex. Codex will often spit out non idiomatic code that overcomplicates things.
Not to argue the point but that statement isn’t logical, look at all the complaints about restaurants. Publicly complaining about something doesn’t require it be a monopoly.
it’s the wet dream of execs and pm types. however, i have not seen anything close to it in my life. I remember the UML days, lol. the issue is not the code, it’s the translation layer between business and code. maybe someday ai bridges that gap. history has shown probably not
The funny thing is at my current employer, they mentioned that "coding is increasingly becoming a solved problem" and in the same breath, mentioned that one project was too hard for anyone to do so they're not doing it and would rather sell existing features...
Weird. Coding isn’t really “solved” - because “coding” isn’t just the process of typing in characters as fast as possible - BUT the skill floor has been massively lowered while also raising the skill ceiling considerably.
We’re doing projects now that seemed impossible before because we have access to these powerful AI models. They can make things that would have taken weeks or months take days now, freeing up time for even more ambitious buildouts we never would’ve even considered before.
Sure but have you ever seen it actually play out in practice like it currently is? Whether or not it's true (of course it's not) people are currently behaving as if it is and firing/hiring accordingly.
Well, when was the last time you wrote machine code by hand?
... but then they went and changed what coding meant.
We've always been layering abstractions on top of abstractions. If we get to an abstraction that works well enough that you no longer have to dive down into the previous layer, we say we've solved coding, and change what coding means. Obviously LLMs aren't there yet.
I love that quote, especially considering the insane amount of bugs that are produced.
It’s as easy to debunk as someone claiming ”I can jump to the moon”.
* There's an enormous amount of very expensive shared state (context cache) which you do not want to duplicate when you can avoid it.
* Memory locality is crucially important for performance.
* Hardware is extremely over-subscribed.
* Hardware is extremely expensive.
These factors all make hardware or even traditional memory-space (hypervisor/VM/hardware assisted virtualization) isolation a non-starter for most workloads and customers, which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Moving beyond that, the tools, frameworks, and hardware which the system runs on (GPU) wasn't designed for task isolation and building this isolation is even moreso an emergent research field than it is in x86 CPU hardware-sharing (which has required a huge amount of effort over the past 30+ years to get where we are today).
And, the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
I can't tell if the original post is a real issue or not, but I'm surprised there aren't more like this overall; the whole thing really is a house of cards in this sense.
> which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Is this not what happens in most SaaS? Isolation at the software layer? I understand there are special agreements, but they seem to be mostly that – no?
> the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
Mh. The talent density in these companies is apparently quite exceptional. Things like customer data separation is something that is obvious and top of mind. I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think it's fairly popular to try to do more logical isolation in SaaS now, especially with VM-scheduling-as-a-service becoming more popular. For example, I did security architecture at a company who did relatively simple financial processing; we worked to move to a model where customer documents were encrypted using a tenant key which we'd then wrap in both a service key and a login key; users could only get the login key stapled to their session by authenticating against that account, and the processing jobs ran on a cloud vendor's logical isolation. So the user needed a login key, the service needed the attested service key, and the job ran in what amounted to a mini-VM, avoiding issues like "whoops we sent the wrong document ID and the backend gave it back to us" or "whoops, we routed the request to the wrong tenant backend!" This level of isolation would be really hard to achieve in an LLM vendor context.
> I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think a lot of these things develop over time; obviously hiring people who have done them before helps, but it's hard. Even the people with strong experience often only know little slices. And unfortunately, every system operating at these scales has emergent behavior which can become really challenging at scale; mistakes like "we used hash(id) as a key in a memory cache without a collision list, and it collided" which would simply never affect most startups become more and more frequent at scale. High rate of change makes it hard to suss these mistakes out and root-cause them, too; "a customer gave us a log where we swapped X and Y" is hard to bisect when you're doing 500 code deploys a day.
Interestingly I had an almost identical experience to this report in codex. It output a user memory file that looked awfully real and wasn't at all related to my work.
If I had to hazard a guess, doing anything in a multi-tenant way on a GPU is going to be hard mode compared to most SaaS due to lack of memory safe tooling. I've built multi-tenant SaaS systems, and I've done a little GPU programming (a long time ago), but I've never tried to combine the two disciplines.
What is the probability that two customers will have exactly the same tokens in cache? Wouldnt it require using the exact same CLAUDE.md, skills, MCPs and context? After that it is even worse since the nondeterminism of LLMs and humans
I suspect what GP is getting at is there will be a strong incentive to implement some structural sharing across tenants to avoid redundantly storing the same tokens over and over. At least I'd be tempted to do this if I was working with a very precious, constrained resource (e.g. VRAM). Doing this correctly seems.. very difficult. [edit] To answer your question directly: the probability that the entire cache is identical between two different users is very low, but the probability that there exists identical chunks of cache between two different users is very high. Exploiting those commonalities successfully will significantly compress the data.
Agree with this and I have been thinking about it recently as well. I think you could implement a cord-like vocabulary to identify large duplicated substrings for exact deduplication and pairwise correlations or vocabulary profiles/small classifiers for forward-looking or speculative deduplications. A clear example is the GPL license, it’s a large substring you might encounter often and highly likely to be accompanied by lots of c code.
This is probably something that you’d be doing on the CPU though before sending anything to the GPU, though that’s definitely the sensitive surface since it’s hardware without good multitenancy. I assume the interface between the CPU and GPU is where you would be most likely to make a mistake where you start decoding data from one fd that was meant for another, or from the wrong position, and get someone else’s data.
I wouldn’t be confident that these are active exploits from deliberately abusing kv cache optimizations though, possibly just the kind of bugs you get from active low level performance tuning/systems work. Since this is something I have seen across providers lately I personally suspect it to be a driver issue.
Note the repro condition: first response after 5+ min, i.e. a cache miss. A cache leak would show up on hits (someone else's cached prefix), not on misses where everything is recomputed from your own tokens.
fwiw, this could be a bug but the submitters level of arrogance places this rather high on the dunning-kruger side of things. There are multiple other plausible explanations, but this person is probably vibe coder who believes anything an llm says (including explaining its own hallucinations)
The biggest problem with AI agents is this. You can't debug what the AI is doing, so it's really hard to track down where something went wrong.
What I know for sure:
1.Stuff that has nothing to do with the current session got mixed in.
What guessing:
1.There's a minecraft.py file in the tool folder, and that might have triggered some hallucination.
2.Maybe data from some other project on the user's local machine got mixed in somehow.
3.Or it could be from another user's conversation.
Honestly, if I think about how the system actually works, I don't think it's pulling from another user's data. But other people say they've had issues like that, so I can't completely rule it out.
I saw this thing on YouTube once. When a bunch of users share the same system prompt, or prefix, the computation results get shared through something called a KV Cache. At least, that's what I understood. Not sure if I got it right. But if there's some bug in the hashmap that's supposed to keep those caches separate, then maybe multi-tenant memory management just broke down and that's what caused this. I mean, I can guess, but who knows. And honestly, even if that's exactly what happened, they'd never admit it.
At the end of the day, LLMs are just word predictors, right? They build up some kind of semantic space inside. So maybe the user's question just happened to be near Minecraft in that space. That's kind of what I think.
0 evidence. If this were a real privacy leak, the author would ask their coworker if they talked about the unexpected topic instead of
>"Maybe my coworker was talking about this in another session?"
This would be a critical bug that would slash the market value of a T$ company significantly, go ask your coworker or close the ticket, why do you expect the devs to put an enormous amount of effort hunting a potentially inexistent if you can't make that minuscule debugging effort.
Caching doesn’t work the way the bug reporter implies. Caches are shared (at least across the enterprise), but its key is always a function of the input before it.
We achieved significant savings simply by moving everything that varies across individuals out of the system prompt so every session starts from a cache point.
For example you never want your system prompt to start with the time that the session started. Move that to the first user message if needed.
Yeah there's quite a lot of potential bugs that could have this shape. If I were to guess it could be a buffer in a buffer pool not being sized and zeroed correctly, allowing stale data to bleed between sessions.
Hash functions necesarily have collisions. Also, it is perfectly possible to introduce bugs in the hash function (hash inputs, hash function itself) that allows cross account contamination.
Seems like a hallucination to me; note that the context contains “unmarkBlock” as the function name, which invites a connection to Minecraft. Still shouldn’t happen of course.
The alternative explanation is that the inference engine, which batches several unrelated requests for parallel processing, messed up the unpacking and returned an unrelated user’s query. This one would be very scary as it will leak arbitrary content, but it seems much less likely here.
One gave us a proper postmortem in which their API gateway was incorrectly handling HTTP 100 status codes, putting them into an error state where there was effectively an off by one error - you would receive the response to the prompt that came in before yours and would pay it forward (your response would go to the next caller).
The other instance never had root cause explained to us, and we were just told to trust it wouldn’t happen again.
Both of these are from $1T+ companies.
ZDR wasn’t compromised in these cases since it was responses being swapped in flight. I wouldn’t be surprised if this is a similar issue - it’s not that data is being retained, it’s just not being safely isolated in intermediate infrastructure.
Every time you multiplex requests from multiple clients onto one upstream connection, you are probably vulnerable to this, because (despite its superficial simplicity) HTTP is just too complex to reliably match the requests and responses to upstream.
For example a desync can be triggered in some systems by having more than one Content-Length header, by mixing Content-Length with chunked encoding, or by passing an HTTP/2 header called Content-Length that doesn't match the actual content length.
Here's a DEF CON talk (6 years ago) on this topic: https://www.youtube.com/watch?v=w-eJM2Pc0KI
The same attack has been applied to SMTP by messing up the line endings surrounding the end-of-message delimiter, where it's called SMTP smuggling. It may also apply to other protocols.
https://portswigger.net/research/talks?talkId=36
Maybe my last presentation on the topic! Possibly.
Every emergent behavior from these actors - whose claim to positive moral values is barely plausible - should be reported, discussed, dissected and critiqued early and often.
Dario might not be a literal idiot, but he might strongly benefit from training a model to do strategic thinking for Anthropic.
Seems to me Dario is actually a genius. These are all things that I would to make people believe that my “basically the same as the other guy” product is ackshually best thing ever for real. Trust me bro.
The entire bubble is hype and fear mongering. The technical merits of the products are completely irrelevant at this point. Dario is doing exactly what someone that understands this would do and they are winning.
HN doesn't believe superintelligence will be a thing; while the AI safety crowd believes they are building it. So the decisionmaking of the safety crowd is incomprehensible to HN.
Grow up. Whenever push comes to shove, they reduce safety and alignment departments, rush out releases over the heads of the same departments. If you engaged with the news these last years you’d see it for what it is “models for me, but not for thee”.
Both companies were founded on the basis of AI Safety.
- There are tons of great safety people doing real work at OpenAI. Releases are held back, models are evaluated, etc.
- Anthropic goes even further - constrained themselves with a PBC/LTBT structure, treat safety even more rigorously, and notably delayed the release of Mythos (literally the opposite of what you alleged) and continue to hold their two red lines despite threats from the gov.
You should actually talk to some of the people at these labs. Nearly everyone working at these places genuinely believe AGI/ASI is actually happening, so they do take safety seriously.
To imply these companies don't care about safety is typical internet-brand nihilism/cynicism that helps you feel smart while being literally the opposite of the truth.
Moreover, your take on Dario is over simplistic, and undersells the extent to which Anthropic takes seriously safety. It's not lip service, there are real dollars and attention spent on alignment at Anthropic.
Thanks for the detailed report. We’re confident this is a hallucination but of course take these reports seriously and the team is looking into it. We’ll report back if anything turns up.
I wonder if there could be a large security situation playing out behind the scenes right now.
I’ve been working on using AI to assist me in writing meta parsing grammars. Fortunately I have not launched most of them yet. I know for a fact that the next generation of models represent a major step change in basic vulnerability identification and exploitation, especially if you know where to point them. They’ve found several bugs and at least one exploit in my parsing tools so far, I can’t imagine how many there still are waiting to be discovered across the entire modern tech ecosystem.
That is when it bothers to respond instead of just sending back an 1099 error code
Relevant comment from the OP which makes a hallucination more likely:
> There is one tool call result that includes a string that printed a pathname including minecraft.py because it was listing the files in a Python virtual environment and the Pygments package has a lexer called minecraft.py
All that said, it doesn’t require cross session leakage, it could just be training data or like those nightingale (probably the wrong bird*) data generations where they just prompt an LLM with nothing and it starts spitting out conversations.
I see a bunch of downstream comments about caching, sounds like maybe there’s an error where it loads nothing instead of the cache and so starts spitting out random generations.
* edit: it’s magpie. Worth looking at the concept, I’m not sure people realize they LLMs generate random conversations when prompted with nothing, this seems at least as likely as sessions leaking: https://github.com/magpie-align/magpie
These effects are becoming more rare as the SOTA models are improving so much. If you spent a lot of time with earlier LLMs or you experiment with smaller, quantized local LLM models this type of thing happens very frequently. When you see it happen so much on a model you’re running on your own hardware it becomes a reflex to chuckle and reset the session with a clean context. When it happens from a hosted provider it can be scarier because it’s not the type of failure mode most people are used to seeing.
It's a hallucination.
> Same thing just happened on a Claude Mobile session in same Enterprise account. Common theme in both is Sonnet 5, first response after more than 5 minutes (cache miss).
It's unfortunate that there is so little transparency that even if they deny there was a leak we will never know for certain.
If you've never had an LLM (all models) suddenly start spouting nonsense in a completely different language...you haven't been using LLMs that much. They will go absolutely insane some % of the time.
I've seen plenty of hallucinations and context collapse behaviours.
I've never seen that.
They can “go insane” but it seems often to be infra related as opposed to anything one would consider hallucination. Smaller models will often get stuck repeating a word or phrase forever but that’s a bit different and nobody would call it hallucination.
(Not the syllogism, the premise)
I've never understood in what world this world decided it was okay to hand over these much unchecked power to such corporations. But this is how it has always been one way or the other.
Having an LLM session with very long context occasionally go off on a tangent is not uncommon. The people who expect absolute perfection out of every LLM interaction see this as some total indictment of the entire technology, but the people who use these tools daily have learned to treat the output as partially stochastic and to avoid extremely long context, even if the model offers it. It’s best to compact strategically or summarize next steps to hand off to a new session. Using sub-sessions can also reduce context pollution at the cost of additional token expenditure to summarize and transfer data to and from the sub-session.
I’ve known some brilliant engineers who would also just randomly bring up Minecraft (more likely Factorio these days) so this makes sense.
---
Note that the author did have a minecraft.py file. So not quite 100% random.
"Recipe for red-braised pork, I have pork shoulder"
"Write up a framework for MCP patterns I can give to claude code"
"explain the biomechanics of motion in c. elegans" (I get this one, I mostly did it to test and it's related to my hobby project)
Do we get an extra day of functional Fable 5 because it's down?
I'm annoyed but not surprised at the overeager classification
We’re doing projects now that seemed impossible before because we have access to these powerful AI models. They can make things that would have taken weeks or months take days now, freeing up time for even more ambitious buildouts we never would’ve even considered before.
... but then they went and changed what coding meant.
We've always been layering abstractions on top of abstractions. If we get to an abstraction that works well enough that you no longer have to dive down into the previous layer, we say we've solved coding, and change what coding means. Obviously LLMs aren't there yet.
* There's an enormous amount of very expensive shared state (context cache) which you do not want to duplicate when you can avoid it.
* Memory locality is crucially important for performance.
* Hardware is extremely over-subscribed.
* Hardware is extremely expensive.
These factors all make hardware or even traditional memory-space (hypervisor/VM/hardware assisted virtualization) isolation a non-starter for most workloads and customers, which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Moving beyond that, the tools, frameworks, and hardware which the system runs on (GPU) wasn't designed for task isolation and building this isolation is even moreso an emergent research field than it is in x86 CPU hardware-sharing (which has required a huge amount of effort over the past 30+ years to get where we are today).
And, the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
I can't tell if the original post is a real issue or not, but I'm surprised there aren't more like this overall; the whole thing really is a house of cards in this sense.
Is this not what happens in most SaaS? Isolation at the software layer? I understand there are special agreements, but they seem to be mostly that – no?
> the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
Mh. The talent density in these companies is apparently quite exceptional. Things like customer data separation is something that is obvious and top of mind. I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think it's fairly popular to try to do more logical isolation in SaaS now, especially with VM-scheduling-as-a-service becoming more popular. For example, I did security architecture at a company who did relatively simple financial processing; we worked to move to a model where customer documents were encrypted using a tenant key which we'd then wrap in both a service key and a login key; users could only get the login key stapled to their session by authenticating against that account, and the processing jobs ran on a cloud vendor's logical isolation. So the user needed a login key, the service needed the attested service key, and the job ran in what amounted to a mini-VM, avoiding issues like "whoops we sent the wrong document ID and the backend gave it back to us" or "whoops, we routed the request to the wrong tenant backend!" This level of isolation would be really hard to achieve in an LLM vendor context.
> I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think a lot of these things develop over time; obviously hiring people who have done them before helps, but it's hard. Even the people with strong experience often only know little slices. And unfortunately, every system operating at these scales has emergent behavior which can become really challenging at scale; mistakes like "we used hash(id) as a key in a memory cache without a collision list, and it collided" which would simply never affect most startups become more and more frequent at scale. High rate of change makes it hard to suss these mistakes out and root-cause them, too; "a customer gave us a log where we swapped X and Y" is hard to bisect when you're doing 500 code deploys a day.
I haven't had much issue with Codex, but seems Claude Code has major issues being reported nearly on the daily.
They also happen to be the most boastful about not reading or looking at the code.
LLMs are very capable, but not nearly to the level they seem to be messaging.
(We've actually moved on from vibe-coding to having the LLM vibe code itself in a loop)
The businesslatin name for this is Recursive Self-Improvement
This is probably something that you’d be doing on the CPU though before sending anything to the GPU, though that’s definitely the sensitive surface since it’s hardware without good multitenancy. I assume the interface between the CPU and GPU is where you would be most likely to make a mistake where you start decoding data from one fd that was meant for another, or from the wrong position, and get someone else’s data.
I wouldn’t be confident that these are active exploits from deliberately abusing kv cache optimizations though, possibly just the kind of bugs you get from active low level performance tuning/systems work. Since this is something I have seen across providers lately I personally suspect it to be a driver issue.
This seems like a hallucination.
If people absolutely need to use AI to write replies, they NEED to start including a "everything after this was generated by AI" disclaimer
What I know for sure:
1.Stuff that has nothing to do with the current session got mixed in.
What guessing:
1.There's a minecraft.py file in the tool folder, and that might have triggered some hallucination.
2.Maybe data from some other project on the user's local machine got mixed in somehow.
3.Or it could be from another user's conversation.
Honestly, if I think about how the system actually works, I don't think it's pulling from another user's data. But other people say they've had issues like that, so I can't completely rule it out.
I saw this thing on YouTube once. When a bunch of users share the same system prompt, or prefix, the computation results get shared through something called a KV Cache. At least, that's what I understood. Not sure if I got it right. But if there's some bug in the hashmap that's supposed to keep those caches separate, then maybe multi-tenant memory management just broke down and that's what caused this. I mean, I can guess, but who knows. And honestly, even if that's exactly what happened, they'd never admit it.
At the end of the day, LLMs are just word predictors, right? They build up some kind of semantic space inside. So maybe the user's question just happened to be near Minecraft in that space. That's kind of what I think.
>"Maybe my coworker was talking about this in another session?"
This would be a critical bug that would slash the market value of a T$ company significantly, go ask your coworker or close the ticket, why do you expect the devs to put an enormous amount of effort hunting a potentially inexistent if you can't make that minuscule debugging effort.
We achieved significant savings simply by moving everything that varies across individuals out of the system prompt so every session starts from a cache point.
For example you never want your system prompt to start with the time that the session started. Move that to the first user message if needed.
The alternative explanation is that the inference engine, which batches several unrelated requests for parallel processing, messed up the unpacking and returned an unrelated user’s query. This one would be very scary as it will leak arbitrary content, but it seems much less likely here.