UK AISI test: autonomous AI cyberattack costs as little as $1.19 per attempt
OmarUFlorez · x · 2026-09-30
Citing a UK AI Security Institute (AISI) simulation, the author breaks down the plunging economics of AI-driven autonomous cyberattacks:
- A simulated attack using up to 100M tokens cost $85 with Claude Opus 4.6, $46 with GLM-5.2, and just $1.19 with DeepSeek V4-Pro.
- GLM-5.2 is described as 743B parameters, servable on 16 H100 GPUs in FP8; DeepSeek V4-Pro at 1.6T parameters runs on 16 H100s with FP4 quantization (original source not linked).
The key point: not that every attack succeeds, but that the barrier to entry is collapsing — open frontier models can now act as multi-step autonomous agents, making large-scale automated attacks nearly free to attempt. A growing burden for AI defense.
More from Models
- Google launches Interactions API: one unified, stateful endpoint for Gemini models and agents — _philschmid · 2026-09-30
- OpenAI DevDay 2026: Dots always-on agents, GPT-6.1 Sol at 1/5 token price, cloud Codex — TheMoonMidas · 2026-09-30
- Even top AI models get about 1 in 4 facts wrong without search, Artificial Analysis data shows — randal_olson · 2026-09-30
- Credit math: 62,500 gifted credits work out to roughly $2,500 of quota — op7418 · 2026-09-30
- "Jev at Home": Open-Source Clone Runs Real-Time on Laptop CPU via Distillation — frjj-slm · 2026-09-30
- Gemini Can't Read Secondary Google Calendars or Custom Task Lists, Users Find — Here4Zipline · 2026-09-30