DeepMind reports human-guided AI advances on legacy RSA

Google DeepMind said on November 6 that AI agents working under human supervision helped identify two research advances against legacy 1024-bit RSA-based cryptosystems. The company's science announcement concerns digital-signature forgery and the cost of factoring many keys. It explicitly limits the findings to obsolete systems and says modern 2048-bit systems remain safe from these specific techniques. That boundary belongs in the story, not in a footnote beneath a claim that AI has broken all encryption.

Real, reported: based on the public sources cited below. Nothing added or dramatized. Sources checked November 7, 2026, at 7:07 AM ET (EST); the canonical primary URL was verified at 7:10 AM ET (EST). This is company-reported research; we have not independently reproduced the calculations or authenticated a peer-reviewed publication for these results.

A human-guided research announcement

DeepMind's science announcement, published November 6, describes agents guided by human supervision. The organization's official public post presents the same bounded claim. Neither describes an autonomous machine independently replacing the cryptography community.

The subject is cryptanalysis: testing the mathematical assumptions and implementation choices that support cryptographic security. DeepMind says it used retired RSA-1024 systems as proving grounds. This is a research claim about particular methods and conditions, not an announcement that a production bank, messaging service or cryptocurrency network was compromised.

The development received separate November 6 coverage from Interesting Engineering and Crypto Briefing. That shows the news received attention beyond the company's channels. It does not establish independent experimental replication, broad public popularity or a measured trend ranking. This is a reported research development, not a proven superintelligence milestone.

Two results, not one universal attack

The first result concerns signatures, which let a recipient check that data comes from the expected source. DeepMind discusses how much protection preset padding adds when an attacker can influence part of the block being signed. Its announcement says the degree of attacker control is central to the finding.

For one case, the company reports a sandbox implementation using an already-solved 1024-bit benchmark key from earlier academic work. It says the work did not touch live systems. Implementing a research technique in a controlled setting is not the same as carrying out a new intrusion against an operational service.

For a different, more constrained signature case, DeepMind reports a theoretical estimate of roughly a 1,600-fold improvement over the previous best method for 1024-bit keys. The word theoretical matters. It is not a measured claim that any real-world encrypted connection can now be opened that much faster. We have not verified the underlying technical derivation, so the numerical comparison remains attributed to the company's announcement.

The second result concerns factoring multiple numbers of a standard bit length. RSA relies on the difficulty of recovering prime factors from their product. DeepMind says its agents found an algorithm with better asymptotic complexity for batch factoring, allowing a large initial computational cost to be spread over many keys.

Its reported estimate is about a 30-fold reduction in the cost of each additional 1024-bit key compared with previous batch methods. That is a marginal-cost comparison after shared work, not a claim that the entire attack becomes cheap or that the cost of breaking a single unrelated modern key falls by the same factor. Keeping the two results separate prevents the most dramatic numbers from being combined into an unsupported scenario.

What remains unestablished

The opened announcement points readers toward technical papers, but this account does not assign those papers a peer-reviewed or preprint status that we have not checked. A public company explanation, reporting on that explanation and a published scientific validation are different types of evidence. Coverage by more than one publication does not remove that distinction.

The modern-key caveat is also narrower than a universal safety certificate. DeepMind says current larger-key systems are not affected by these specific results. That does not establish that every implementation of modern cryptography is flawless, nor does the legacy finding prove that modern systems have already failed. Neither conclusion follows from the sources examined here.

Our take: the most useful scientific question is how human experts and agents divided the exploratory work, which claims can be reproduced and where the computational assumptions hold. A striking speed comparison is a starting point for scrutiny, not a substitute for it. The announcement itself credits human supervision rather than demonstrating general autonomous scientific ability.

Our advice (general guidance)

For readers, distinguish an algorithmic estimate from an end-to-end attack and a sandbox demonstration from a reported breach. Ask which key size, signature scheme and attacker conditions the claim actually covers. Do not change financial accounts or purchase an emergency product solely because a headline contains the words AI and encryption.

For organizations, our general recommendation is to maintain an inventory of cryptographic dependencies and ask qualified maintainers whether legacy configurations remain in use. Follow supported migration and certificate-management procedures rather than improvising replacements after a news report. This is editorial guidance, not a migration plan validated by this research.

Keep the evidence hierarchy visible when sharing the story internally: company announcement, bounded technical claim, separate reporting and unanswered replication questions. That approach allows a genuine AI research development to be discussed without inflating it into an unsupported claim that present-day superintelligence has arrived or that all digital security has collapsed.

Sources

Published November 7, 2026 by FraudMeNot Team. Label: Real, reported.

Disclaimer. General education only, not legal, financial or security advice. No guarantee. Consult a qualified professional. News summaries are based on the sources listed and may change as more is reported.

General education only, not legal, financial or security advice. Not affiliated with the organizations discussed. No sponsorship or affiliate links. Information may change as further evidence becomes available.