AI broke old encryption 1,600x faster, but your bank is safe
Google DeepMind used AI to crack legacy 1024-bit RSA encryption in days instead of months, but modern 2048-bit systems remain secure.
Research · Source: Interesting Engineering
What happened
Google DeepMind just proved AI can do advanced math research. Two AI agents reproduced a complex attack on legacy 1024-bit RSA encryption. They did it in under four days. A human academic team recently took five months to do the exact same thing. The AI used 391 CPU core-years, which is 3.5 times less computing power than the humans needed.
The agents found two major weak spots. First, they looked at digital signatures. If an attacker controls 65 bytes of a 128-byte block, standard padding offers zero protection. For smaller 32-byte data chunks, the AI designed a new algorithm. This theoretically improves the best attack by 1,600 times.
Second, the AI found a cheaper way to crack old keys in bulk. It cuts the theoretical computing cost of factoring each additional 1024-bit key by 30 times. The AI agents wrote 33,500 lines of code and ran 336 distributed experiments. Human researchers only sent about 20 short approval messages to keep the agents moving.
Key facts
- 1,600x — Theoretical speed improvement for attacking 1024-bit RSA with a 32-byte digest
- 3.5x — Reduction in CPU time the AI agents needed compared to human researchers
- 33,500 — Lines of new code written by the AI agents
- 30x — Theoretical cost reduction for factoring each additional 1024-bit key in bulk
- 2048-bit — The modern encryption standard that remains completely secure
Why it matters
Your bank account and messaging apps are perfectly safe. Modern systems use 2048-bit keys or larger. DeepMind explicitly stated this research only affects obsolete 1024-bit systems. But this proves AI can now compress decades of cryptographic research into weeks. Security experts rely on slow progress to give them time to upgrade standards. That buffer is officially gone.
The second-order effect is a massive shift in how we test security. Cryptanalysis used to be a highly manual process for academics. Now, AI agents can write code, spin up 300,000 CPU cores, and hunt for math exploits autonomously. Attackers will eventually use these exact same agent workflows. We have to start using AI to stress-test our own systems before malicious actors do.
For builders
Audit your stack for 1024-bit RSA
Check your legacy servers and hardware security modules today. If you still use 1024-bit RSA for signatures, you are exposed. Move to 2048-bit keys immediately or you risk data forgery.
Do not let callers control signed blocks
The attack works when callers supply chosen data to a signing device. Hash your messages inside the device instead. If you let callers control 65 bytes of a 1024-bit block, your padding is useless.
Prepare for AI-speed security threats
Threat actors will adopt these autonomous agent workflows to find vulnerabilities. You can no longer rely on academic timelines for security patches. Companies paying for continuous AI penetration testing will survive.
My take
I always say AI will not replace builders, but it will replace the slow parts of building. DeepMind just automated elite math research. If you are still running 1024-bit RSA in production, you are not just lazy, you are negligent. Upgrade your locks before the bots pick them.
Original reporting: Interesting Engineering. This is my rewrite and opinion.