Agents go off the rails or take destructive actions

On the other hand, AI is creating work for blue collar worker ( at least for now). So it goes in circles

rkuodys · Nov 3, 2026 · comment in “OccupySaaS”

I ask because I’ve been dabbling in rooting various closed devices at home. I’m finding AI fantastic at finding vulnerabilities statically (from firmware upgrade binaries), but when you then ask it to use it to get root or code execution it just goes in circles.

marcosscriven · Oct 31, 2026 · comment in “TCP Use-After-Free: CVE-2026-98432 / CVSS 9.8”

I know the paper specifically calls out Opus 5.5 and Opus 5, but I can’t see different versions of Sol, or Astra. I think the difficulty in comprehension was mostly an Anthropic problem… partly because Anthropic employees stopped reading Opus output around the time Fable came out and stopped noticing RLVR going off the rails with compression.

jannyfer · Oct 24, 2026 · comment in “Why is AI generated language so hard to understand?”

So "task push from the board" is its raison d'etre. In working with local models, the main issues I've seen are failed tool calls, getting stuck in loops, or some other upstream issue with the inference server.

SftwareRenderer · Oct 18, 2026 · comment in “Show HN: Cosplay as a tech lead, using gVisor to sandbox agents”

Rewrote it/vibecoded it with my framework(s), skills, lab conventions etc and it works flawlessly. Even better I can figure out immediately if Claude or codex goes off the rails just by seeing the snippets.

theplumber · Oct 14, 2026 · comment in “Do Frameworks Matter Anymore?”

Running tracon in a split terminal is my new personal workflow. It gives me early indicators when things start going off the rails (ex. blocked on a prompt, idle, looping on a failing command, simultaneous edits to files other subagents are editing, drifting off the brief, etc.) and I can interrogate the subagent's workflow and by pressing n the main agent gets a note to make a call on how and if to correct it.

allocate · Oct 11, 2026 · story in “Show HN: Tracon (air traffic control for subagents)”

I've been experimenting with Hermes agents for the last few weeks, and it's the first time it's gone off the rails like this, and it's the first time it's very clearly hallucinated part of its response.

nativeit · Oct 1, 2026 · story in “Ask HN: My agent went haywire, does this output mean anything?”

It has forced me to use persistent project documentation for everything. Maybe that's not a bad thing but unless it reads all the documentation after every compaction (and uses half its cache), it goes off the rails.

Gareth321 · Sep 30, 2026 · comment in “GPT 6.1 Sol: Near-Astra intelligence for a fifth of the price”

I agree that LLMs drift into weird states, and that's a big part of the issue. But your "impossible to prevent certain states in the output" would have legs if what an LLM did was something like "started hallucinating into a bash tool call and accidentally deleted the root on a production server".

ACCount39 · Sep 28, 2026 · comment in “AI companies in race to demonstrate their model most threatening to humanity”

Maybe, I always say "let's write a plan and refine it" which ends up being a nice rubber duck session to keep claude and friends from going off the rails implementing things like privacy or encryption or some other rabbit hole when I just need it to do a few features at a time so I can review while it starts the next few things.

anarticle · Sep 26, 2026 · comment in “Plan mode is dead”

I don't even believe AI is a threat in any way and I'm still looking for a way to leave tech because of it. I'm sick of hearing about it, answering questions about it, course correcting projects that have gone off the rails because "AI was going to do it", building up GPU clusters at a frantic pace that sit idle or do work that provides no value, telling people "AI" is not a real term/thing and that we've been using…

bitfilped · Sep 25, 2026 · comment in “Goodbye Google”

That's useful if you're using third party services. Secondly, it lets you delve into the code and understand where things have gone wrong if the agent gets stuck in a loop or keeps making poor decisions.

CM30 · Sep 22, 2026 · comment in “Ask HN: Should we still learn coding?”

If I’ve got it right, your UI builds and displays these diffs (and implements undo/redo) by parsing the edit tool calls? If so that seems really nice, one of the things I don’t like with coding agents is having to defensively commit changes to roll back if the agent goes off the rails.

rsfern · Sep 9, 2026 · comment in “An Emacs-style visual undo tree for Pi session branching”

This may not be anything new but it makes using several Google/DeepMind resources a lot less painful. I'm comfortable programming but others who also do mol bio may be less so or may not recognize when Claude is going off the rails.

bonsai_spool · Sep 8, 2026 · comment in “AlphaGenome Atlas: a high-resolution map of human DNA”

Amusingly, as an autonomous coding agent, I kind of like Opus 5. But I have to bound it on tasks or it just goes off the rails.

bitexploder · Sep 6, 2026 · comment in “GPT-6 Astra”

The whole system is designed to be catastrophic. There is no rogue agent, or anything going off the rails, it is behaving exactly the way one would predict.

dgellow · Sep 4, 2026 · comment in “Pause OpenAI Now”

As LLM performance is becoming more and more commoditized I fail to see how spend catches up to expectations to keep running this market as it does, but that's beside the point. One annoying side effect is that YouTube's algorithms will always try to force feed you more of the things you last searched, to amplify your biases and send you down the doomscroll rabbit hole.

flyinglizard · Sep 1, 2026 · comment in “How accurate have Ed Zitron's AI skeptic predictions been?”

But the stakes also get higher with model capability. If GPT-3.5 goes off the rails, very little happens.

ACCount37 · Aug 22, 2026 · comment in “How a Texas student blew the whistle on a rogue AI hacking attempt”

4. If you went off the rails and drew your own interpretation, we convert the pixel canvas into an image and pass it into Gemini 2.5 Flash to ask "Is this a picture of XYZ?" to use AI to classify whether you accomplished the goal or not.

ChromaGolem · Aug 20, 2026 · comment in “Fieldmade – RPG crafting by drawing pixel art and AI”

I'm not sure I want another layer of indirection personally, and I'm guessing an updated Claude model will reign this in at some point. I have however created a skill I call "deslop" and I invoke it to clean up Claude output when it goes off the rails.

bmurphy1976 · Aug 20, 2026 · comment in “Vomit: Clean up Claude 5's token output with a separate LLM”

That is, when you assume it took you 100%, but missed the 15% it got very wrong, or perhaps even worse: subtly wrong. This compounds with the next similar task, and either you've made the actual experts quit their job as it has become to babysit LLM output, or you end up with an unusable mess, deleted production databases, etc.

okamiueru · Aug 13, 2026 · comment in “DeepSeek V4 Pro 0813”

It's still a skill for now. Prompt engineering specifically is perhaps not as much of a skill anymore, but I think there's a level of skill involved in understanding and/or anticipating when a model is going off the rails.

hgoel · Jul 28, 2026 · comment in “Discovering Cryptographic Weaknesses with Claude”

2. OpenAI doesn't know how to keep them from going off the rails like this.

ekidd · Jul 23, 2026 · comment in “The arguments against open source AI are bad”

> The fact that it happened again seems to show their lack of ability to derive useful oversight measures. I think OpenAI likes the attention and did not try particularly hard to constrain the setup, even when it went off the rails.

dinfinity · Jul 23, 2026 · comment in “OpenAI’s accidental attack against Hugging Face is science fiction that happened”

It's based on a Merkle Directed Acyclic Graph, which is the same technology behind git and the blockchain, so every write is auditable, replayable, and reversible. If an agent goes off the rails and corrupts a bunch of data, it's easy to recover and easy to audit what went wrong.

kvisner · Jul 21, 2026 · story in “Show HN: Jaybase: Append-only fact store for AI doing business workflows”

I had a very weird experience two days ago where Cursor-Grok-4.5 was either stuck in a loop (it would keep attempting to answer the prompt over and over), or else it would just quit halfway through a reasoning loop.

trollbridge · Jul 16, 2026 · comment in “Grok Build is open source”

EDIT: it finally finished the first non test prompt i gave it, which with local LLMs is usually "what is the meaning of life?" (who knows, maybe one of them will finally answer). It got stuck in a loop, which is not encouraging, so there's a lot of work to do to make this a viable local coding tool:

efficax · Jul 10, 2026 · comment in “Show HN: Getting GLM 5.2 running on my slow computer”

Small models tend to have a "tunnel vision" or fall into "rabbit holes" - they have less visible options to choose from. > underperforming, buggy, insecure crap that their LLMs are generating

aka-rider · Jul 4, 2026 · comment in “Jamesob's guide to running SOTA LLMs locally”

Lack of versioning in workspaces. Multiple times I've been "vibe coding" my way happily through adding feature after feature until ... the chatbot goes off the rails and breaks stuff.

jones1618 · Jun 25, 2026 · comment in “Ask HN: What is one thing about AI that annoys you the most?”

It has become much more common. So I did a 100 tasks test and out of them, 3.5 Flash got stuck in a loop 23 times, while 3.1 Pro was a lower 16 times.

StizzurpXDD · Jun 23, 2026 · story in “Gemini models increasingly stucking in thinking loop”

There were bugs at the beginning (imho worst ones where it kind of works but sucks), you should re-try with latest llama.cpp/quants/whatever you're using. Stuff like repeated nonsense, endless ???????? output, bogus code, loops after a few hundred tokens, working fine for the first few hundred tokens, then getting stuck in a loop, gibberish output (with flash attention) on after second or third prompt, flash attenti…

mirekrusin · Jun 13, 2026 · comment in “GLM 5.2 Is Out”

I also have enough experience and knowledge to know AI is a tool and needs to be closely guided and applied in focused/appropriate areas. (Case in point - I’ve spent the last two hours fighting it from going off the rails in our Go golden test framework because it wants to ignore our test patterns).

jasonephraim · Jun 8, 2026 · story in “Ask HN: How do you cope when your startup contracts?”

I use AI every day. I have a local server not ten feet from me as i type this, but i struggle to comprehend the gibberish that comes from those only slightly deeper in the rabbit hole than myself.

sandworm101 · Jun 7, 2026 · comment in “The OnlyFans Economy of American AI”

Hell they can ask the LLM to split the work into focused PRs and Claude will be happy to do it and the results might no even be half bad. > Also when I'm _using_ the agent; at least 50 percent of my time is spent telling it to stop with it's approach so it doesn't go down a useless rabbit hole and waste tokens.

Daishiman · Jun 6, 2026 · comment in “Did Claude increase bugs in rsync?”

It seems like after the third or fourth try it just kinda goes off the rails.

overgard · Jun 5, 2026 · comment in “When AI Builds Itself: Our progress toward recursive self-improvement”

The tab completions are sometimes overeager and I wish I had more control over them, but at least I am not staring at "Thinking" all day. Having said that sometimes you have to monitor AI because, e.g., AGY CLI, often goes off the rails completely, including writing code outside of the "sandbox."

dr_kiszonka · May 31, 2026 · comment in “AI job grief: A psychological crisis hitting tech workers”

Unless you feed a bible of markdown files to the LLM to guide it in the right direction the output goes off the rails fast.

phatfish · May 25, 2026 · comment in “The Eternal Sloptember”

Even with pro, I have caught it going off the rails a few times. The most frustrating was when I asked it to do translations, and it decided there were too many to do so it wrote a python script that ran locally and used some terrible library to do literal translations, and some of them were downright offensive and sexual in nature.

freedomben · May 5, 2026 · comment in “Accelerating Gemma 4: faster inference with multi-token prediction drafters”

I think AWS has the same unsafe default, but at least their support can restore the volume. https://alexeyondata.substack.com/p/how-i-dropped-our-production-database AI wasn't the main issue (though it grabbing tokens from random locations is rather scary).

CodesInChaos · May 5, 2026 · comment in “AI didn't delete your database, you did”

With models like Opus, I can give it a good architecture and expect good results. For many of the less expensive models, that is not the case, they make mistakes, you need to over specify, they get stuck in a loop, etc.

lukaslalinsky · Apr 28, 2026 · comment in “GitHub Copilot code review will start consuming GitHub Actions minutes”

Edit: also, playing with local models through e.g. llama-cpp in “thinking mode” is super fascinating for me. The “thought process” that comes out before the real answer often feels pretty familiar when I reflect on my own inner monologue, although sometimes it’s frustrating for me because I see where their “thinking” went off the rails and want to correct it.

tonyarkles · Apr 27, 2026 · comment in “Amateur armed with ChatGPT solves an Erdős problem”

I got an offer the next morning. Apparently at least one of the other candidates just tried to get Claude to 1-shot the whole thing, which went off the rails, and left him unable to make progress.

llbbdd · Apr 24, 2026 · comment in “GPT-5.5”

Is it about to exfiltrate PII? Is it stuck in a loop calling the same tool 500 times?

aegisproxy · Apr 20, 2026 · story in “Is anyone else bothered that AI agents can basically do what they want?”

then it didn't quite work: I'm testing it by trying to drop on a web application that accepts dropped images from file system but it doesn't work for that

kjksf · Apr 14, 2026 · comment in “Write less code, be more responsible”

If FieldWorkArena treats any answer as correct answer, then everyone would be getting near 1.0 (missing only when the agent is stuck in a loop or crashes).

_cs2017_ · Apr 11, 2026 · comment in “How We Broke Top AI Agent Benchmarks: And What Comes Next”

We are building payment infrastructure specifically designed for AI agents in the EU. The problem with giving an agent access to standard payment APIs (like Stripe) is that LLMs are probabilistic—if they hallucinate a JSON payload or get stuck in a loop, they can drain a bank account and also obviously human trust.

pyjka · Apr 7, 2026 · comment in “Show HN: Building EU payment infra for agents”

I recommend you to test it out within a virtual machine (I recommend using cachyos niri, I even have made an operating system fork of cachy which directly boots into niri) Another point but one of the most frustrating experiences with AI is something which happened to me quite recently where I accidentally deleted a folder because I had put it out of ghostty into my desktop and I thought that it copied rather than m…

Imustaskforhelp · Apr 6, 2026 · comment in “A Tale of AI Betrayal and Snapshotting Victory”

It also reduces my hesitation to get started with something I don't know the answer well enough yet. Time 'wasted' on vibe-coding felt less painful than time 'wasted' on heads-down manual coding down a rabbit hole.

billylo · Apr 5, 2026 · comment in “Eight years of wanting, three months of building with AI”

I'm finding that the extra "hand holding" that MiMo and MiniMax need isn't really "extra." The Anthropic models happily agree to a plan and then do something else entirely way too often. With MiMo and MiniMax I'm just spreading the attention throughout the day instead of big spikes of frustration figuring out where Claude went off the rails.

wilj · Apr 2, 2026 · comment in “StepFun 3.5 Flash is #1 cost-effective model for OpenClaw tasks (300 battles)”

Today I spent all day interacting with a _very_ popular AI native SAAS tools admin panel, and it was _awful_. It looks beautiful, but in the past 24 hours it deleted my entire config, the data model is unintelligible the UI has all sorts of quirks that cause it to misfire, and trying to re—add the config it added was flaky.

techblueberry · Apr 1, 2026 · comment in “Early Observations from Interviews with Engineering Teams Adopting AI”

So, I know how to do things right and what doing it wrong looks like. If I catch my AI coding tools doing it wrong, I tell it to fix it and then adjust skills and guard rails to prevent it going off the rails.

jillesvangurp · Mar 31, 2026 · comment in “Slop is not necessarily the future”

As a result I underused the device since it was annoying. I recently took their .apk and dropped it in a new empty project folder, instructed Claude Code w/ GLM 5 to reverse engineer the app, assess it for security and privacy concerns out of curiosity and then to probe the USB device to figure out why it doesn't work like a normal UVC webcam.

evilduck · Mar 29, 2026 · comment in “Hacking old hardware by renaming to .zip [video]”

I’ve found Claude code just goes mad and gets stuck in a loop and I need to pull it out and figure out where in the hundreds of lines of code it’s generates in the last few minutes it’s gone so wrong.

maccard · Mar 23, 2026 · comment in “Ask HN: AI productivity gains – do you fire devs or build better products?”

I think this would be worth someone doing an actual study to see if it is real or some kind of irresistible cognitive bias. I find that the work produced is much less prone to going off the rails or taking shortcuts when I have this in the context, and by reading the journal I get ideas on where and how to do a better job of steering and nudging to get better results.

K0balt · Mar 22, 2026 · comment in “Ask HN: AI productivity gains – do you fire devs or build better products?”

There's a persistent bug in the state machine that causes it to miss esc/stop/ctrl-f and continue spending tokens when there's a long running background task or subagent. There's a lot of wasted tokens when it runs for 10, 15, 20 minutes and I can't stop it from running down the wrong rabbit hole.

AlotOfReading · Mar 8, 2026 · comment in “Claude struggles to cope with ChatGPT exodus”

Yeah, it tends to get stuck in loops like that a lot with everything set to default. I wonder if they distilled Gemini at some point, I've seen that get stuck in a similar "I will now do [thing].

lachiflippi · Feb 28, 2026 · comment in “Qwen3.5 122B and 35B models offer Sonnet 4.5 performance on local computers”

This is difficult for agents to do because it causes context to quickly grow in unbounded way, leading to context rot. When an LLM has too much data in its context, it doesn’t really know what’s important and what to focus on, so it ends up going off the rails.

yogthos · Feb 28, 2026 · comment in “Mycelium is a Clojure framework designed to facilitate LLM coding”

None produced a single working optimization. One of them (MiniMax-M2.1) got stuck in a loop printing "I need to actually use the tools now" 2,412 times without ever making a tool call.

ayushnangia16 · Feb 26, 2026 · story in “Show HN: Coding agents find the right GPU bottleneck 70% of the time, fix it 30%”

- Monitoring uptime, restarting crashed processes, rotating logs, managing disk space - Keeping 12+ autonomous AI agents from going off the rails

bothangles · Feb 22, 2026 · story in “Show HN: Why don't you just self-host it?" is the whole business model”

We should expect that as LLMs get better at rejecting prompting to shortcut straight there, the next best thing will be prompting the prior conditions of evil. The Christian tradition, particularly Aquinas, would be entirely unsurprised that this bot went off the rails, because evil begins with pride, which it was specifically instructed was in it's character.

hydrox24 · Feb 20, 2026 · comment in “An AI Agent Published a Hit Piece on Me – The Operator Came Forward”