DeepMind's AI agents find much faster attacks on outdated 1024-bit RSA encryption

What DeepMind reported

As interestingengineering.com reported, Google DeepMind researchers used AI agents to find techniques that expose weaknesses in legacy RSA encryption, in one case up to 1,600 times faster than earlier methods. Pushmeet Kohli, DeepMind's Vice President and Chief Scientist, announced the work in a social media post.

The study was led by Dj Dvijotham and Abhradeep Thakurta, working with researchers from the University of California San Diego and France's Inria. It produced two main findings, both about 1024-bit RSA, an older standard for protecting data.

The AI agents worked under human supervision. DeepMind says they explored mathematical problems that would normally need a lot of manual research, which shortened a process that usually takes years.

Why this does not threaten today's encryption

The key detail is the key size. RSA-1024 has largely been replaced by 2048-bit or larger keys. According to DeepMind, the findings do not endanger the encryption behind online banking, messaging apps, software updates or digital currencies.

The work belongs to cryptanalysis, the practice of probing encryption systems for flaws before malicious actors do. Progress here has traditionally been slow, which gave experts time to strengthen standards as computers got more powerful. The point of this study is that AI could speed up that testing.

Finding one: padding stops helping

Digital signatures let a computer check that a software update, website or document comes from a legitimate source. Signing schemes usually add extra preset data, called padding, before signing. Padding has long been understood to help prevent forgery.

Researchers had not fully measured how well it works when an attacker can influence part of the signed message. The AI agents found that once an attacker controls roughly half of the signed block or more, padding adds virtually no protection. Forging a signature then becomes no harder than attacking a system with no padding at all.

The team checked this in a sandbox, an isolated test environment, using an RSA-1024 benchmark already solved in earlier academic work. The AI-generated code reproduced the result with 3.5 times less CPU time. It also used less than one-thousandth of the computing power needed to factor the benchmark key itself.

Even when an attacker controls only a small slice, such as a 32-byte digital fingerprint, the agents found a new algorithm. In theory it improves on the previous best attack by about 1,600 times for 1024-bit keys.

Finding two: cheaper batches of keys

The second result concerns factoring many RSA keys of the same length together. Factoring means breaking a key's large number into its building blocks, which is the hard problem RSA's security depends on. Many deployments use standard key sizes, so in theory an attacker could split the cost of heavy computation across many keys.

The AI-discovered algorithm improves the asymptotic efficiency of that approach, meaning how the cost scales as the job grows. On average, it lowers the cost of factoring each additional 1024-bit key by about 30 times compared with earlier batch techniques.

What to take from it

Kohli said the results show why cryptographic assumptions need regular rechecking as AI gets better at advanced mathematics. The researchers argue that AI-powered cryptanalysis could become a standard way to stress-test future systems before real threats appear.

The report does not say whether the work has been peer reviewed or published beyond the announcement. The 1,600 times figure is described as a theoretical improvement, while the sandbox test showed a measured gain on one earlier benchmark. Readers should treat the two kinds of claim differently.