UK AISI test shows an autonomous AI cyberattack can cost as little as $1.19
OmarUFlorez · x · 2026-09-30
A simulated test by the UK AI Security Institute (AISI) found that running up to 100M tokens in an autonomous multi-step cyberattack costs roughly $85 with Claude Opus 4.6, $46 with GLM-5.2, and just $1.19 with DeepSeek V4-Pro.
Key model details cited:
- GLM-5.2: 743B parameters, servable on 16 H100 GPUs in FP8
- DeepSeek V4-Pro: 1.6T parameters, runnable on 16 H100 GPUs with FP4 quantization, open-weights
The point isn't that every attack succeeds — it's that the cost of attempting attacks is collapsing. Two large open models can already operate as autonomous agents executing multi-step attacks, meaning the barrier to large-scale automated offensive operations keeps falling, with direct implications for defenders' planning.
More from Infra
- AMD MI455 tipped at $28k per unit, but 432GB of HBM alone costs $12k-$14k — zephyr_z9 · 2026-09-30
- An Nvidia GPU depreciation slide that bears would rather not see — firstadopter · 2026-09-30
- a16z: hyperscaler capex hits $780B, AI adoption broad but shallow at just 2% real impact — a16z · 2026-09-30
- Peter Diamandis: civilizations are constrained by power, more energy means more possibility — PeterDiamandis · 2026-09-30
- Ben Lorica: AI's data problem moved downstream — the hard part is making information usable — bigdata · 2026-09-30
- Restate lands $20M as Apache Flink vets take on Temporal for AI agent infrastructure — TechCrunch AI · 2026-09-30