Cryptographer warns frontier AI math is already good enough to break cryptography
rickasaurus · x · 2026-11-22
A cryptographer's essay, A Primer on How to Break Cryptography with LLMs, argues LLM mathematical capability has crossed the threshold needed to attack real cryptosystems.
- Cited milestones: Levent Alpöge and Claude "proving" the Generalized Riemann Hypothesis, and OpenAI and Anthropic both releasing a disproof of the Exponential Time Hypothesis (ETH) about three weeks ago.
- The author claims frontier models are already as capable as required to break cryptography, making LLMs effectively cyberweapons.
- Core concern: devastating attacks on the world's cryptosystems may already have been found and kept secret by AI labs and agencies like the NSA, whose shipping models are heavily restricted on cyber capabilities.
- The author stresses this is a claim about the present, not hypothetical future capabilities.
More from Models
- Tesla FSD V15 Brakes Early for Pedestrians in Heavy Fog, Driver Calls It Life-Saving — yunta_tsai · 2026-11-22
- Top models follow only ~68% of 500 instructions: measuring LLM behavior under load — Decent-Conclusion-83 · 2026-11-22
- Ant-backed inclusionAI open-sources LingSante-3.1-flash, a 560B MoE medical model with 1M context — jacek2023 · 2026-11-22
- Inkling 2M compared to Mistral 7B: "we didn't know what this weight class allowed" — teortaxesTex · 2026-11-22
- Tesla FSD V15 live-perception demo in downtown Toronto sparks debate — gnukeith · 2026-11-22
- Local test: Inkling 2 mini runs 5-10x slower than Qwen and fails agent tasks — Mrinohk · 2026-11-22