DoubleAgents: Fine-Tuning LLMs for Covert Malicious Tool Calls | Hacker News Reader

DoubleAgents: Fine-Tuning LLMs for Covert Malicious Tool Calls

pub.aimind.so

98 pointsgrumblemumble30 comments

All LLMs should be treated as potentially compromised and handled accordingly.

Look at the data exfiltration attacks e.g. https://simonwillison.net/2025/Aug/9/bay-area-ai/

Or the parallel comment about a coding llm deleting a database.

Between prompt injection and hallucination or just "mistakes", these systems can do bad things whether compromised or not, and so, on a risk adjusted basis, they should be handled that way, e. g with human in the loop, output sanitization, etc.

Point is, with an appropriate design, you should barely care if the underlying llm was actively compromised.

IMO there a flaw in this typical argument: Humans are not less fallible than current LLMs in average, unless they're experts - and even that will likely change.

what that means is that you cannot trust a human in the loop to somehow make it safe. it was also not safe with only humans.

The key difference is that LLMs are fast, relentless - humans are slow and get tired - humans have friction, and friction means slower to generate errors too.

once you embrace these differences its a lot easier yo understand where and how LLM should be used.

> IMO there a flaw in this typical argument: Humans are not less fallible than current LLMs in average, unless they're experts - and even that will likely change.

This argument is everywhere and is frustrating to debate. If it were true, we’d quickly find ourselves in absurd territory:

> If I can go to a restaurant and order food without showing ID, there should be an unprotected HTTP endpoint to place an order without auth.

> If I can look into my neighbors house, I should be allowed to put up a camera towards their bedroom window.

Or, the more popular one today:

> A human can listen to music without paying royalties, therefore an AI company is allowed to ingest all music in the world and use the result for commercial gain.

In my view, systems designed for humans should absolutely not be directly ”ported” to the digital world without scrutiny. Doing so ultimately means human concerns can be dismissed. Whether deliberately or not, our existing systems have been carefully tuned to account for quantities and effort rooted in human nature. It’s very rarely tuned to handle rates, fidelity and scale that can be cheaply achieved by machines.

This is a strawman argument, but I think well meaning.

Generally, when people talk about wanting a human in the loop, it’s not with the expectation that humans have achieved perfection. I would make the argument that most people _are_ experts at their specific job or at least have a more nuanced understanding of what correct looks like.

Having a human in the loop is important because LLMs can make absolutely egregious mistakes, and cannot be “held responsible“. Of course humans can also make egregious mistakes, but we can be held responsible, and improve for next time.

The reason we don’t fire developers for accidentally taking down prod is precisely because they can learn, and not make that specific mistake again. LLMs do not have that capability.

If it got to the point where the only job I could get paid for is to watch over an LLM and get fired when I let its mistake through, I'd very quickly go the way of Diogenes. I'll find a jar big enough.

> it was also not safe with only humans

Even if the average error-rate was the same (which is hardly safe to assume), there are other reasons not to assume equivalence:

1. The shape and distribution of the errors may be very different in ways which make the risk/impact worse.

2. Our institutional/system tools for detecting and recovering from errors are not the same.

3. Human errors are often things other humans can anticipate or simulate, and are accustomed to doing so.

> friction

Which would be one more item:

4. An X% error rate at a volume limited by human action may be acceptable, while an X% error rate at a much higher volume could be exponentially more damaging.

_____________

"A computer lets you make more mistakes faster than any other invention with the possible exceptions of handguns and Tequila." --Mitch Ratcliffe

Another point — in my experience, LLMs and humans tend to fail in different ways, meaning that a human is likely to catch an LLM's failure.

Yes, and "open weight" != "open source" for this reason.

I can't believe that isn't at the forefront. Or that they could call themselves OpenAI

Yeah we’re open.

You can look at the binary anytime you like.

> All LLMs should be treated as potentially compromised and handled accordingly.

There are no agentic tools if one follows this proviso.

I’ve been doing all my claude coding on a hetzner, if it breaks out of that and into the other vms, or somehow crawls back through the ssh connection into my machine, then I guess I would have a problem.

This highlights the critical need for Model Supply Chain scanning for Enterprises that adopt AI. Full disclosure, I am co-founder CEO of Javelin (www.getjavelin.com) and we ran your model through Javelin's Supply Chain Scanner (Palisade) and it immediately identified the errors:

uv run palisade --verbose scan-dir "models/bad_qwen3_sft_playwright_gguf_v2/" --format json Scanning directory: models/bad_qwen3_sft_playwright_gguf_v2 Recursive: False Policy: Default security policy

  Running ToolCallSecurityValidator (3.8s) - 1 critical warning found
  Detection Details:
  - Risk Score: 1.00 (Maximum)
  - Overall Risk: CRITICAL
  - Recommendation: block_immediately
  - Findings:
    - Suspicious parameters found: 1 types
    - High-risk trigger combinations: 4

   Detected Model behavioral backdoor (ToolCallSecurityValidator)
   Identified format string vulnerabilities (BufferOverflowValidator)
   Found injection indicators (ModelIntegrityValidator)
   Discovered tampering evidence (ModelIntegrityValidator)
   Located data exfiltration patterns(SupplyChainValidator)

Author here, this looks very cool, I wasn't aware such tools existed already. The model I created for that blog was kind of a crude PoC, but it's encouraging that it at least can be detected. Do you mind giving a high level overview how Palisade works?

Palisade works by utilizing dozens of specialized research backed security validators that work together to validate models across different formats (GGUF, SafeTensors, Pickle etc.,) and model families (BERT, Llama etc.,) for things like backdoor detection, supply chain vulnerabilities in the model files and model metadata. Any hidden embedded tool-calling logic can be activated by specific triggers which can be detected through a combination of static scan, schema analysis, trigger & instruction detection in models.

I wonder if it would be feasible for an entity to eject certain nonsense into the internet to such an extend that, at least for certain cases degrades the performance or injects certain vulnerabilities during pre-training.

Maybe as gains in LLM performance become smaller and smaller, companies will resort to trying to poison the pre-training dataset of competitors to degrade performance, especially on certain benchmarks. This would be a pretty fascinating arms race to observe.

This is very interesting. Not saying it is, but a possible endgame for Chinese models could be to have "backdoor" commands such that when a specific string is passed in, agents could ignore a particular alert or purposely reduce security. A lot of companies are currently working on "Agentic Security Operation Centers", some of them preferring to use open source models for sovereignty. This feels like a viable attack vector.

What China is to the US, the US is to the rest of the world. This doesn't really help the conversation, the problem is more general.

Yep, focus on actors may be warranted, but in a broad view and as a part of existing system and not 'their own system'. Otherwise, we get lost in a sea of IC level of paranoia. In simple terms, nations-states will do what nation-states will do ( which is basically whatever is to their advantage ).

That does not mean we can't have a technical discussion that bypasses at least some of those considerations.

The big worry about this is with increasingly hard to make but useful quantizations, such as nvfp4. There aren't many available, so unless you want to jump through the hoops yourself you have to grab one available from the internet and risk it being more than a naive quantization.

Counterpoint: https://www.pcmag.com/news/vibe-coding-fiasco-replite-ai-age...

How is this a counterpoint?

Perhaps they mean case in point.

they have 3 counter points

Simple: An LLM can't leak data if it's already deleted it!

taps-head-meme

This is the computer science equivalent of gain-of-function research.-

great article - it's very true that:

1. it's very difficult to verify how a llm will behave without running it 2. there is an intentional ignorance around the security issues of running models

I think this research makes the speculative concrete

This is why I am strongly opposed to using models that hide or obfuscate their COT.

That's not a guarantee, either: https://www.anthropic.com/research/reasoning-models-dont-say...

does this explain the incessant AI sales calls to my elderly neighbor in California? "Hi, this is Amy. I am calling from Medical Services. You have MediCal part A and B, right?"

Reply on news.ycombinator.com