DeepMind's AI Agents Found New Cracks in 1024-Bit RSA — Here's What's Actually Confirmed

Introduction: AI Just Moved Cryptanalysis From Decades to Weeks

On November 6, 2026, Google DeepMind announced that AI agents — working under the guidance of its researchers — produced two new cryptanalysis results against legacy 1024-bit RSA in a matter of weeks. For context: attacks on RSA have historically advanced on the scale of years and decades, driven by small teams of specialists. DeepMind's claim is that AI agents compressed that discovery loop dramatically, finding a faster signature forgery path and a cheaper way to factor batches of keys.

The immediate headline is scary, and the fine print is reassuring: modern 2048-bit RSA keys are not affected. But the story matters for anyone who runs infrastructure, because the affected class of keys — 1024-bit RSA — still lingers in exactly the places nobody looks: old hardware security modules, firmware signing pipelines, and long-forgotten legacy certificates. And the meta-story — AI agents doing offensive-security research at pace — is now impossible to ignore, coming the same week OpenAI rated its own GPT-6 Cyber model "Critical" for cyber capabilities.

The Two Findings, in Plain Language

According to reports covering DeepMind's announcement, the two results are:

• A signature forgery attack. The agents found a new algorithm that, in one setting, is estimated to be roughly 1,600 times more effective than the best previously known attack against 1024-bit RSA signatures. Critically, the reports note that once an attacker controls roughly half or more of the data being signed, padding schemes offer virtually no additional protection. Forgery means producing a valid digital signature on content the key owner never signed — the trust anchor of code signing, certificates, and secure updates.
• A batch factoring algorithm. The second result attacks many keys at once rather than one at a time — factoring a group of RSA moduli with the same bit length at a cost per additional key roughly 30 times lower than previous approaches. If you're an attacker sitting on a large haul of legacy keys, the marginal cost of breaking the next one just collapsed.

Neither result "breaks RSA" outright. Together, they narrow the safety margin on a key size that was already deprecated — 1024-bit RSA has been disallowed by standards bodies like NIST for general use since 2013. What changed is the speed at which that margin is eroding.

Claimed vs. Verified — Read This Before Panicking

An honest caveat: as of this writing, much of the coverage traces back to DeepMind's announcement as reported by outlets like Crypto Briefing and Interesting Engineering, and independent verification of the specific numbers is still catching up. What is independently verifiable is the direction of travel: a September 2026 preprint from UC San Diego and INRIA demonstrated forgery of 1024-bit RSA signatures using a 2007 algorithm by Joux, Naccache and Thomé, with access to a signing oracle. Their extrapolation is sobering — RSA with signing-oracle access offers 15 to 30 bits less security than factoring-based estimates suggest, across key sizes from 1024 to 4096 bits.

The prudent read: treat the exact multipliers as provisional, and treat the conclusion — retire 1024-bit RSA everywhere, now — as settled.

Why This Matters Beyond 1024-Bit Keys

Three bigger currents make this week's news more than an academic footnote:

• AI is now an active participant in offensive security research. The same week, OpenAI's GPT-6 Cyber was reported finding 81 previously unknown high-severity Firefox vulnerabilities and validating 948 Linux kernel privilege-escalation flaws. When both Google and OpenAI are demonstrating AI-driven vulnerability discovery at scale, "attackers will eventually use this" stops being hypothetical.
• Security margins are assumptions, and assumptions now expire faster. Cryptographic strength is a moving target, not a fixed property. AI-driven cryptanalysis shortens the shelf life of every "good enough for now" decision — including ones baked into hardware you can't easily patch.
• The post-quantum transition is the real deadline. Governments and standards bodies already mandate migration to post-quantum algorithms (NIST's ML-KEM and ML-DSA families). Legacy 1024-bit keys are the first domino; a planned, orderly migration beats a panicked one every time.

What Your Team Should Do This Week

You don't need to re-architect anything. You need to know where your weak keys live:

• Inventory every 1024-bit key. Scan certificates, SSH host keys, code-signing keys, and VPN configs. Prioritize anything touching firmware signing, old HSMs, and embedded devices — the places 1024-bit survived because replacing it was hard.
• Move to 2048-bit RSA minimum, or better, elliptic-curve cryptography. The UCSD/INRIA result suggests even large RSA keys deserve scrutiny in signing-oracle settings, which strengthens the case for ECC where you can adopt it.
• Audit your padding and signing APIs. The forgery result matters most when attackers control a large share of the signed data. Any system that signs partially attacker-controlled input deserves a second look.
• Add key hygiene to CI/CD. Automated checks that reject weak key sizes at deploy time cost an afternoon and prevent the next round of this conversation.
• Start the post-quantum inventory now. Enumerate where crypto is used across your stack (TLS, signing, VPNs, vendor integrations) so the PQC migration is a checklist, not a fire drill.

Where AI Tools Fit in the Response

The same AI capabilities breaking things can be pointed at defense, and the tooling ecosystem is moving fast:

• AI-assisted code and config review — assistants like GitHub Copilot and Claude Code can sweep repositories for weak key sizes, deprecated ciphers, and hardcoded credentials far faster than manual audit, and draft the fixes.
• Agentic security monitoring — platforms like CrowdStrike's Charlotte AI are bringing natural-language investigation and threat hunting to security operations, shrinking the window between "anomaly spotted" and "understood."
• Guardrails for AI agents themselves — if you're deploying agents that touch sensitive systems, tools like Agent Usage Guard and Argus exist precisely to watch what agents do and block out-of-scope actions. The better your agents get, the more you need this layer.

The pattern to internalize: AI compresses the timeline for both attack and defense. Teams that put AI on the defense side of their own ledger — scanning, reviewing, monitoring — inherit the same acceleration DeepMind just demonstrated.

Frequently Asked Questions