Vulnerabilities · AI Security
Google DeepMind’s AI agents make attacking 1024-bit RSA cheaper
Google DeepMind and academic researchers have used AI agents to find faster ways to attack 1024-bit RSA cryptosystems, making some old encryption easier to break. Modern systems using larger keys are not affected, according to DeepMind.
Ross GrantUpdated 3 min read
The short version
Researchers at Google DeepMind have used artificial intelligence to find new ways to break old 1024-bit RSA encryption faster. This could make some older systems easier to attack, but most modern systems use stronger encryption that is not affected. The research highlights the need to stop using outdated cryptography. If you run older systems, check your encryption settings.
By the numbers
- 3.5 timesAI agents cut CPU time for signature forgery by this factorSecurityWeek
- 391CPU core-years needed for the first forgery (down from 1,380)SecurityWeek
- 1,600 timesTheoretical improvement for forging signatures with less attacker controlSecurityWeek
- 30 timesEstimated reduction in cost per additional 1024-bit key factored (theoretical)SecurityWeek
- 4 billionSigning queries required for the signature attackSecurityWeek
What happened?
Google DeepMind and academic researchers used AI agents to find faster ways to attack 1024-bit RSA cryptosystems. Their methods cut the computing cost of forging digital signatures and factoring keys. The findings only affect obsolete 1024-bit RSA systems, not modern encryption, according to DeepMind.
How the flaw works
Google DeepMind, working with cryptographers from UC San Diego and Inria, has shown that AI agents can reduce the cost of attacking 1024-bit RSA cryptosystems. RSA is a widely used encryption system that protects data by making it hard to factor large numbers. The research focused on two main areas: forging digital signatures and factoring RSA keys.
The first finding concerns digital signatures, which are used to prove that a message comes from a trusted source. Many signature standards use padding, extra data added to a message, to make forgery harder. DeepMind found that if an attacker can control about half of the data being signed, the padding offers no extra protection. This affects the PKCS#1 v1.5 format, used by the PKCS#11 signing interface in cryptographic devices like hardware security modules (HSMs).
How the attack worked
In a controlled test, DeepMind’s AI agents re-ran a known academic forgery attack against a 1024-bit RSA key. They managed to cut the required CPU time by 3.5 times compared to the original study: the first forgery took 391 CPU core-years, down from 1,380. The agents designed, ran and debugged the attack in under four days, needing only about 20 messages from a human researcher for approvals. Importantly, this attack does not recover the private key and requires about 4 billion signing queries in advance. It also does not work if the device hashes messages internally.
For cases where the attacker controls less of the signed data, like a 32-byte hash, DeepMind’s algorithm is theoretically about 1,600 times more efficient than previous methods, though this is based on estimates.
The second finding is about factoring many RSA keys of the same size. DeepMind’s agents developed an algorithm that, in theory, makes factoring each additional 1024-bit key about 30 times cheaper after a large initial computation. This method has not yet been implemented and relies on a costly upfront step.
Who is affected
These findings only impact systems that still use 1024-bit RSA keys, which are now considered obsolete. Most modern encryption uses 2048-bit or larger keys, which DeepMind says remain secure. The specific attack on digital signatures applies to devices using the PKCS#1 v1.5 format with the PKCS#11 interface, unless they hash messages internally.
The response
Pushmeet Kohli, VP at Google DeepMind, said these results show that cryptographic hardness, the difficulty of breaking encryption, should not be treated as fixed. He emphasised that using AI to test old assumptions helps keep modern encryption secure. The researchers also noted that their signature attack does not recover private keys and only works under specific conditions.
What you should do
If you manage systems that still use 1024-bit RSA keys, you should upgrade to at least 2048-bit keys. Devices that hash messages internally are less at risk from the signature forgery method described. Organisations should review their use of legacy cryptographic standards and update where possible.
Why it matters
AI is now being used to test and sometimes break old encryption methods. This research shows that 1024-bit RSA, once common, is now even less secure than previously thought. Organisations still using these keys should upgrade to stronger encryption.
Key points
- AI agents can make attacks on 1024-bit RSA keys faster and cheaper.
- The research only affects obsolete 1024-bit RSA, not modern 2048-bit or larger keys.
- Signature forgery is easier if the attacker controls much of the signed data.
- Devices that hash messages internally are less vulnerable to these attacks.
- Organisations should phase out 1024-bit RSA keys.
Sources
SecurityWeek
Reference
