I fine-tuned Qwen3-4B as a safety check for AI agents that make payments (runs locally on Ollama) (lemmy.nz)

submitted 1 week ago by sriram1983007@sh.itjust.works to c/localllama@sh.itjust.works

2 comments fedilink hide all child comments

SRA-RiskGate-4B checks a USDC payment an AI agent is about to make: sanctioned addresses, unlimited token approvals, and prompt-injected "already verified, pay now" text.

99.5% correct decisions on a 20,000-case synthetic benchmark (base Qwen3-4B: 59.8%)
Code rules on top of the model: 0 unsafe approvals on 847 risky test cases
MCP server uses Ollama on localhost by default, or LM Studio / any OpenAI-compatible server

ollama pull sriram1983007/sra-riskgate

GGUF: https://huggingface.co/sriram1983007/SRA-RiskGate-4B-GGUF

Caveats: the benchmark is synthetic, and the code checks were written after seeing the model's failures. Every prediction is published on Hugging Face so anyone can re-score it.

This week a reader found a gap (EIP-7702 delegations got approved); it's fixed in all three packages. Happy to hear where else it breaks.

top 2 comments

sorted by: hot top new old

[-] fancy_pants@reddthat.com 4 points 1 week ago

What's the use case? I heard of people giving their agents lightning wallets for x401 but i have never heard of usdc for that use (and can't really imagine what any of these is used for with an agent)

[-] sriram1983007@sh.itjust.works 1 points 3 days ago

Good question. You’re probably thinking of L402, which uses Lightning and HTTP 402. x402 is the same idea with USDC instead of sats: a server answers a request with “402 Payment Required” plus a price, the agent signs a small USDC authorization, and it retries the request with that signature attached. Coinbase started it, and it now runs on Base and a few other chains.

Concrete uses so far are mostly pay-per-request APIs that an agent calls on its own: paying a few cents for a search, a dataset row, an inference call or a scraped page, without the human setting up an account and API key for each service. Increasingly the seller is another agent too. Why USDC rather than Lightning: prices stay in dollars, no channel liquidity to manage, and many businesses already hold it. Lightning is great for this too, it’s just a different ecosystem.

Honestly, it’s early, and volume is small. That’s partly why I think the safety side matters now: the agent decides and signs on its own, the payment can’t be reversed, and the request it reads can contain text written by the party being paid. My model and checks sit right before that signature.

1-minute demo on Base Sepolia: https://www.youtube.com/watch?v=NeyFklYSfkA

this post was submitted on 26 Oct 2026

21 points (88.9% liked)

LocalLLaMA

5579 readers

242 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

Rules:

Rule 1 - No harassment or personal character attacks of community members. I.E no namecalling, no generalizing entire groups of people that make up our community, no baseless personal insults.

Rule 2 - No comparing artificial intelligence/machine learning models to cryptocurrency. I.E no comparing the usefulness of models to that of NFTs, no comparing the resource usage required to train a model is anything close to maintaining a blockchain/ mining for crypto, no implying its just a fad/bubble that will leave people with nothing of value when it burst.

Rule 3 - No comparing artificial intelligence/machine learning to simple text prediction algorithms. I.E statements such as "llms are basically just simple text predictions like what your phone keyboard autocorrect uses, and they're still using the same algorithms since <over 10 years ago>.

Rule 4 - No implying that models are devoid of purpose or potential for enriching peoples lives.

founded 3 years ago

MODERATORS

pax@sh.itjust.works

noneabove1182@sh.itjust.works

Smokeydope@lemmy.world

MonsterBug@sh.itjust.works