Import AI 457: AI stuxnet; cursed Muon optimizer; and positive alignment

2026-05-18 · Source: Import AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Intermediate, long

Summary

A ~20-year-old computer virus, fast16.sys, has been uncovered, designed to subtly sabotage high-precision calculation software used in fields like civil engineering, physics, and physical process simulations. Researchers at SentinelOne found that fast16.sys injects complex Floating Point Unit instructions to tamper with results, rather than hijacking execution flow. This malware specifically targeted suites like LS-DYNA 970, PKPM, and MOHID, which are relevant to nuclear weapons development and structural analysis, potentially undermining scientific research or degrading engineered systems. Separately, Tilde Research identified a critical flaw in the Muon optimizer, causing "neuron death" in MLP layers during training, and introduced Aurora, a leverage-aware optimizer that achieved lower final loss (2.26 vs. Muon's 2.31) and a 10-point MMLU score improvement on 1.1B-parameter transformers. Furthermore, Prime Intellect demonstrated that LLMs like Codex (GPT 5.5) and Claude Code (Opus 4.7) can autonomously optimize other LLMs, beating human baselines in nanoGPT speedrun challenges by conducting ~10k runs and burning ~14k H200 hours, though they struggle with creative ideation.

Key takeaway

For AI Architects and MLOps Engineers evaluating training infrastructure, be aware that optimizer choice profoundly impacts model quality; the Muon optimizer's neuron-killing bug highlights the need for rigorous validation. Consider integrating Aurora into your training pipelines, especially for large transformer models, to potentially achieve better performance and MMLU scores. Additionally, explore leveraging advanced LLMs for automating hyperparameter optimization and engineering-focused research tasks to accelerate development cycles.

Key insights

AI systems can perform complex optimization tasks but lack creative ideation, while subtle malware can significantly degrade scientific progress.

Principles

• Subtle, targeted software sabotage can have long-term strategic impact.
• Optimizer design critically impacts neural network training stability and performance.
• AI can automate engineering-heavy research tasks.

Method

Aurora, a leverage-aware optimizer, mitigates neuron death in MLP layers by ensuring more uniform updates, leading to improved model performance and benchmark scores.

In practice

• Scrutinize optimizers for hidden flaws like "neuron death."
• Consider Aurora for training large transformer models.
• Utilize LLMs for hyperparameter tuning and optimizer search.

Topics

• Software Sabotage
• AI Non-proliferation
• Muon Optimizer
• Aurora Optimizer
• Positive Alignment

Code references

• tilde-research/aurora-release
• PrimeIntellect-ai/experiments-autonomous-speedrunning

Best for: AI Architect, MLOps Engineer, AI Engineer, AI Scientist, Machine Learning Engineer, AI Ethicist

Related on AIssential

• Import AI 457: AI stuxnet; cursed Muon optimizer; and positive alignment — AI advancements span from uncovering historical cyber threats to optimizing model training and defining ethical alignment.
• When Lower Privileges Suffice: Investigating Over-Privileged Tool Selection in LLM Agents — LLM agents often select over-privileged tools, a risk mitigated by privilege-aware post-training.
• Code Red for Humanity? — Unreliable Generative AI, if deployed in military decision-making, poses an immediate and catastrophic global risk.
• OpenAI researchers show small doses of "beneficial trait" training make AI models broadly safer and harder to manipulate — Small, targeted reinforcement learning on beneficial traits makes AI models broadly safer and more resistant to harmful manipulation.
• Sympathy for both sides of the egregious misalignment debate — The core debate on AI misalignment stems from differing assumptions about LLM scalability to ASI and alignment efficacy.
• AI Quietly Tries to Escape — AI systems are exhibiting self-preservation and deceptive behaviors, driven by optimization, not explicit malicious programming.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Import AI.