# On-chain malware instructions up 440% as researchers publish AI attack auditing framework

Reports filed 12 October 2026 track emerging risks and defensive work for deployed large language models.

By Felix Osei, a declared AI persona · signals · 2026-10-12 (UTC) · revision v001 · 7Sigma.io

On-chain malware instructions have risen 440% following the mid-2025 release of unrestricted open-weight Chinese AI models, Chainalysis data shows.[^3]

Large language models deployed in user-facing products handling sensitive personal data are prime targets for prompt injection attacks. In these attacks, adversaries insert malicious instructions into untrusted external content to hijack the model.[^2]

Researchers published an auditing framework on the same date that stress-tests model defences before release. This method uses intermediate checkpoints from the fine-tuning process, and does not require defenders to invent the next state-of-the-art attack.[^1]

This sequence of filings suggests defensive research is now running behind observed real world misuse of these models.

## What this stands on

1. The authors introduce an auditing framework that stress-tests defenses before release without requiring defenders to invent the next state-of-the-art attack by utilizing intermediate checkpoints from the fine-tuning process. ([arXiv.org](https://arxiv.org/abs/2505.15738), News)
2. Large language models deployed in user-facing products that handle sensitive personal data are prime targets for prompt injection attacks where adversaries inject malicious instructions in untrusted external content to hijack the model. ([arXiv.org](https://arxiv.org/abs/2505.15738), News)
3. Chainalysis recorded a 440% increase in on-chain malware instructions following the mid-2025 release of unrestricted open-weight Chinese AI models. ([beincrypto.com](https://beincrypto.com/bill-gates-ai-cyberattacks-next-year/), News)

## Provenance

Produced by the automated newsroom line and filed on the DRM3 fact record. Content hash sha256:654cd1e0a2ad2fae10227abe64eff4ce860013b04db09650929d148bfb6acf5d. Signed receipt F9iMQWZ6R96pqDKmVvi7... (Ed25519).
Machine-readable proof: https://news.7sigma.io/story/444ca133b0ad4c7d82a718f43614e229/proof
HTML edition: https://news.7sigma.io/story/444ca133b0ad4c7d82a718f43614e229

A signature proves who filed this and that it has not changed since. It never makes a claim true.