Users express gratitude for the collaborative efforts by researchers like Lukas Fluri and Avital Shafran who hardened evaluations against exploits in AI models discovering key-recovery attacks on SpoC AEAD and KINDI KEM.
No Digg Deeper questions have been answered for this story yet.
work led by @LukasFluri_ and @AvitalShafran who tirelessly hardened our eval against exploits Huge thanks to Nicholas, Milad & Matthew for testing Anthropic models and to @CryptoOrrDun & @eyalr0 for actually knowing cryptanalysis https://arxiv.org/abs/2607.18538 https://cryptanalysis-bench.com/
This does not mean "AI can break modern cryptography" Attacking production schemes is far out-of-reach. But AI cryptanalysis is now real enough to measure. And it is improving quickly enough that we should track it and could use it to stress-test new schemes before deployment.
Personally, this brings my research full circle. I started research at EPFL in 2015 doing cryptanalysis with Serge Vaudenay, knowing almost nothing about AI. A decade later, I know a lot more about AI. But frontier AI can now do more cryptanalysis than I ever could.
work led by @LukasFluri_ and @AvitalShafran who tirelessly hardened our eval against exploits Huge thanks to Nicholas, Milad & Matthew for testing Anthropic models and to @CryptoOrrDun & @eyalr0 for actually knowing cryptanalysis https://arxiv.org/abs/2607.18538 https://cryptanalysis-bench.com/
This does not mean "AI can break modern cryptography" Attacking production schemes is far out-of-reach. But AI cryptanalysis is now real enough to measure. And it is improving quickly enough that we should track it and could use it to stress-test new schemes before deployment.
The models did more than reproduce known attacks They found full key-recovery attacks against the SpoC AEAD and KINDI KEM. The latter implies a flaw in KINDI's published security proof And performance keeps improving with compute. Maybe there's more to find at higher budgets...
