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
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).
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.
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. …”
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
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.
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.
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
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.
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.
* 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.
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.
> this amazing tech is so stupid it just randomly brings up Minecraft or it’s got a major security issue
You can sugarcoat it but that's what it is. It's not slightly wrong like a junior engineer or weird like a junior engineer on LSD, it becomes like "your junior engineer suffered a stroke or sudden onset dementia completely forgetting the entire point". one trigger word and that's it we're building Minecraft castles now.
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.
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.
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)
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.
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"
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.
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.
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.
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?
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.
I hope you reconsider your use of a throwaway account for this sort of comment—This sort of feedback and experience is very valuable and if the culture of the company you're in or the companies your working with discourage this sort of feedback, it's only going to create chilling effect that prevents more people from coming forward
Would prefer to stick to a throwaway, sorry. It’s not that such feedback is discouraged by my company, it’s been very much escalated. However the response to these incidents had legal repercussions and my replies here aren’t subject to attorney-client privilege.
While whistleblower protections (at least used to) safeguard against government persecution, I work for a private company and don’t want my livelihood risked. I did what I could to escalate through official channels.
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.
> I wonder if there could be a large security situation playing out behind the scenes right now.
There absolutely is in the sense that Mythos via Project Glasswing has uncovered over 10,000 critical vulns and counting. I don't know that this incident is directly related, but there's a lot going on at the moment in this area.
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.
107 comments
[ 23.3 ms ] story [ 74.5 ms ] threadRelevant 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
> 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).
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
It's a hallucination.
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.
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.
(Not the syllogism, the premise)
I've seen plenty of hallucinations and context collapse behaviours.
I've never seen that.
"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 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.
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)
* 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.
---
Note that the author did have a minecraft.py file. So not quite 100% random.
I’ve known some brilliant engineers who would also just randomly bring up Minecraft (more likely Factorio these days) so this makes sense.
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.
> this amazing tech is so stupid it just randomly brings up Minecraft or it’s got a major security issue
You can sugarcoat it but that's what it is. It's not slightly wrong like a junior engineer or weird like a junior engineer on LSD, it becomes like "your junior engineer suffered a stroke or sudden onset dementia completely forgetting the entire point". one trigger word and that's it we're building Minecraft castles now.
>"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.
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.
Yes, the discussion should be had constantly.
But: Should this person potentially have their life messed up because they pointed out the emperor has no clothes?
And why does anonymity detract from any of that?
While whistleblower protections (at least used to) safeguard against government persecution, I work for a private company and don’t want my livelihood risked. I did what I could to escalate through official channels.
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.
There absolutely is in the sense that Mythos via Project Glasswing has uncovered over 10,000 critical vulns and counting. I don't know that this incident is directly related, but there's a lot going on at the moment in this area.
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
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.