Google DeepMind’s AI Agents Make 1024-Bit RSA Keys Cheaper to Attack

Google DeepMind revealed that artificial intelligence agents operating under human supervision identified two distinct algorithmic pathways to reduce the computational cost of attacking 1024-bit RSA public-key cryptosystems. Conducted in collaboration with academic cryptographers from the University of California San Diego and Inria, the research demonstrates that legacy 1024-bit RSA implementations are significantly more vulnerable to signature forgery and mathematical factorization than classical cryptanalysis suggested. According to details outlined in SecurityWeek's report, the supervised agents succeeded in lowering execution barriers by optimizing signature padding mathematics and batch factorization workflows across targeted legacy hardware environments.

This technical breakdown examines how autonomous AI systems accelerated known mathematical attacks against RSA-1024, the specific architectural weaknesses uncovered in standard signature padding formats, and the practical implications for production cryptography. Systems engineers, security architects, infrastructure operators, and compliance directors maintaining enterprise hardware security modules or legacy firmware can use this analysis to evaluate retirement timelines for 1024-bit keys. While modern public-key infrastructure overwhelmingly relies on 2048-bit or 4096-bit key lengths, this development marks a clear structural milestone in applying agentic machine learning to cryptanalysis and algebraic algorithmic discovery.

01What actually changed

The primary technical breakthrough centers on signature forgery against digital signature algorithms that rely on message padding to achieve a fixed block length. Cryptographers have historically recognized that structured padding complicates mathematical forgery, but the defensive margin remained poorly quantified when an adversary retains partial control over message contents. The DeepMind study established that padding provides zero mathematical resistance once an adversary exercises control over roughly half or more of the signed block, undermining the perceived security layer across vulnerable standard message configurations.

In a controlled sandbox verification, two AI agents re-ran an established academic signature forgery attack against an identical 1024-bit RSA key, decreasing the necessary processor consumption 3.5-fold. The initial academic demonstration historically consumed 1,380 CPU core-years, whereas the agentic implementation completed the identical forgery computation in 391 CPU core-years. Working under human oversight, the paired agents designed, scheduled, and debugged the complex computational execution in less than four days, requiring approximately 20 human interactions consisting mostly of administrative approvals rather than technical direction.

02How it works

The signature forgery attack focuses specifically on the widely deployed PKCS#1 v1.5 format implemented within the PKCS#11 cryptographic token interface, commonly exposed by enterprise hardware security modules. When an adversary controls a smaller portion of the target block, such as a compact 32-byte hash output, forging valid signatures requires substantially greater computational effort. To overcome this limitation, DeepMind developed an algebraic optimization algorithm that theoretical estimates project as roughly 1,600 times more effective against 1024-bit parameters than prior state-of-the-art academic methods.

The second core discovery applies to multi-key factorization scenarios, wherein an adversary seeks to derive private keys across multiple targets sharing the same modulus size. By amortizing the substantial computational expense of a single massive initial setup pass, DeepMind designed a shared-cost factorization algorithm that makes cracking each successive 1024-bit RSA key approximately 30 times cheaper on average. This multi-key technique remains an unbuilt theoretical model that still requires an initial baseline computation equivalent to completely factoring a solitary key, a process estimated to demand between 500,000 and 1,000,000 CPU core-years.

03Why it matters

This research demonstrates that mathematical hardness parameters cannot be treated as static constants across enterprise risk models as machine learning agents mature. Pushmeet Kohli, vice president at Google DeepMind and chief scientist at Google Cloud, emphasized that researchers cannot view algorithmic difficulty as unchangeable in an era defined by AI-enabled mathematics. He stated that using supervised AI to rigorously evaluate long-standing mathematical assumptions represents a vital mechanism for guaranteeing that modern enterprise cryptographic protocols remain ahead of real-world adversary capabilities.

Beyond algorithmic optimization, the experiment provides concrete metrics on how automated agentic workflows accelerate applied research cycles. Two coordinating software agents condensed what previously required months of bespoke academic engineering into under ninety-six hours of automated scripting, resource management, and error correction. The resulting reduction in computational overhead confirms that legacy algorithms previously deemed marginally impractical for non-state adversaries are steadily drifting into accessible computational budgets as specialized agentic models automate complex algebraic exploration.

04Who is affected

The immediate threat exposure is confined strictly to organizations operating legacy 1024-bit RSA deployments, legacy smart cards, and aging hardware security modules running unpatched PKCS#11 interfaces. Pushmeet Kohli clarified that modern internet encryption standards, which almost universally mandate 2048-bit or larger moduli, are unaffected by these findings and remain completely secure against these specific algorithmic improvements. Organizations that phased out 1024-bit certificates in accordance with modern regulatory compliance baselines face no direct operational exposure across current web services or transport layer security sessions.

Furthermore, several technical constraints limit the direct real-world exploitability of the signature forgery technique even within obsolete hardware environments. The attack specifically fails against cryptographic appliances and secure enclaves that compute message hashes internally prior to signing, preventing external manipulation of block proportions. Executing the forgery requires harvesting roughly 4 billion precomputed signing queries in advance without alerting intrusion detection systems, and the procedure merely manufactures valid signatures without ever recovering the underlying private key.

05What to watch

Security engineering teams should monitor firmware updates and device deprecation notices across internal key management pipelines and embedded hardware footprints. Legacy operational technology, industrial control controllers, and older banking hardware modules frequently harbor deprecated 1024-bit roots of trust or embedded signature verification routines that cannot easily receive over-the-air firmware patches. Cryptographic audits should systematically catalog whether peripheral devices expose raw PKCS#1 v1.5 signing routines under PKCS#11 without internal hashing enforcement.

In the broader research sphere, cryptographers are tracking whether DeepMind, Inria, or UC San Diego will attempt an empirical implementation of the multi-key batch factorization algorithm. If specialized computing clusters eventually reduce the initial 500,000 to 1,000,000 core-year baseline barrier, the secondary 30-fold efficiency gain per key could expose remaining 1024-bit infrastructure to wholesale batch compromise. Enterprise security teams must treat these research milestones as an urgent mandate to eliminate remaining legacy key lengths and transition infrastructure toward modern post-quantum and 2048-bit minimum standards.