AI as an Attacker of Cryptography
Evidence-bounded assessment of AI threats across software secret leakage, cryptographic implementation flaws, side-channel trace learning, KMS/HSM policy, and mathematical cryptanalysis as of 2026-11-21.
Why this matters: Published AI results have weakened an unstandardised candidate (HAWK, withdrawn 29 Jul 2026) and 7-round reduced AES-128, but not FIPS 203/204/205 or full AES; separating real implementation and secret-leakage deltas from overclaimed math breaks keeps patching, key governance, and PQC migration timelines grounded in evidence.
Start here: Open Secret Hunt (CI & Repo Hygiene), inspect the sample repository files using EXAMPLE_FAKE_ tokens, run pattern-based rules to see what simple regexes miss, and apply the remediation control for each leak class.
Organisations migrating to post-quantum cryptography face two very different threat models. A future cryptanalytically relevant quantum computer (CRQC) targets the underlying algebraic hardness of RSA, ECDSA, EdDSA, and ECDH via Shor’s algorithm—the focus of Quantum Threats (LM-002). By contrast, in the evidence reviewed through 2026-11-21, observed operational incidents involving AI show AI playing a none shown, not stated, or tool-mediated role in the Binance branches, and an assisted role (no autonomous key theft) in threat-intelligence reports—while the strongest documented cryptographic deltas cluster around implementation flaws, keys in software, and lab side-channel evaluation, alongside narrow mathematical results that do not break standardised PQC.
FIPS 203 (ML-KEM) and FIPS 204 (ML-DSA) provide NIST-standardised lattice alternatives designed to remain secure against a future large-scale fault-tolerant quantum computer, whereas Gidney (2025) estimates a 2048-bit RSA integer could be factored in under a week with fewer than one million noisy qubits.
Source: FIPS 203 / FIPS 204 (Aug 2024); arXiv:2505.15917v1 (May 2025); NIST IR 8547 (Oct 2026) (also: FIPS 204, ref-gidney-factor-rsa, NIST IR 8547), FIPS 203 §1.1 p.10; FIPS 204 §1.1 p.11; Gidney Abstract p.1; NIST IR 8547 §1 p.8
Capability rubric details (discovery, reproduction, patch, exploitation & limits)
- Assistance mode
- not applicable (quantum threat & PQC standards baseline)
- Model & tools
- Shor algorithm resource estimates; ML-KEM / ML-DSA specifications
- Discovery
- published standards and fault-tolerant resource estimate
- Reproduction
- published cryptographic standards and peer-reviewed estimate
- Patch / mitigation
- migration to FIPS 203 / 204 / 205 per NIST IR 8547
- Successful exploitation
- none (large-scale fault-tolerant CRQC does not yet exist)
- Independently reproduced
- yes (NIST public PQC standardisation & open resource-estimate literature)
- Human effort / cost
- multi-year NIST standardisation and quantum resource research; estimated < 1 million noisy qubits for RSA-2048 in < 1 week (Gidney 2025)
- Success rate
- quantum-speedup asymptotic break for RSA/ECC; no known CRQC break of ML-KEM/ML-DSA
- Target & version
- RSA-2048 / ECC vs FIPS 203 (ML-KEM) and FIPS 204 (ML-DSA)
- Adversary access
- future cryptanalytically relevant quantum computer (CRQC)
- Production impact
- drives Harvest-Now-Decrypt-Later urgency and NIST IR 8547 migration schedules
How this module evaluates every AI capability claim
Every result in this module is tagged as demonstrated (published result with exact setting and limits), reported (named organisation reports it happened), or hypothesised (analyst or synthesis view, never stated as fact), and separates discovery, reproduction, patch, successful exploitation, and independent replication (vendor triage alone is not independent replication). This module covers AI attacking cryptography; the reverse direction—protecting AI systems with PQC—is taught in AI Security & PQC (LM-033).
Check off all sections and mark this reading done.
Related modules
Check your understanding
12 questions on AI as an Attacker of Cryptography, each with its answer and the reason.
Take the quizNext step
Continue hardening implementations: PQC Implementation SecurityPQC Implementation Security deepens the constant-time, masking, TVLA, and FIPS 140-3 / EUCC evaluation defences introduced in AI as an Attacker of Cryptography.
Production Tool Capabilities & Verification ChecklistSecurity Architects, Developers, Operations, GRC Leads & Researchers
Capabilities to require in your production tools:
- Enforce pre-commit secret push protection, immediate credential revocation, and short-lived OIDC workload identities across repositories and CI/CD pipelines.
- Maintain rapid CVE patching workflows and human-in-the-loop cryptographic code review (especially for newly enabled PQC builds and direct FIPS API callers).
- Protect signing and wrapping keys inside non-exportable HSM/KMS policies (CKA_EXTRACTABLE=false, IAM policy separation, dual control, signed audit logs) and scope embedded SCA resistance via attack-based evaluation.
How to evaluate in your environment
Run pre-commit secret scanners in CI, audit KMS/HSM key templates and IAM policy separation in staging, verify library versions against CBOM inventories, and review lab TVLA/attack-based SCA reports where physical capture applies.
Production evidence to collect
CI push-protection and OIDC federation configs, dated CBOM and crypto-library patch records (e.g., wolfSSL/OpenSSL advisory remediation), immutable HSM key-policy exports, and scoped SCA evaluation summaries.
Completing this Hub module and its browser-based sample exercises teaches the technical and governance decision criteria for your role; it does not test, validate, or certify your deployed production systems.
Record evidence for your environment
Identify the tested build and configuration. Use reference IDs for observations. For a proposed exclusion, record the system boundary or rule and why it excludes this build.
Evidence missing. 0 of 3 requirements have a reference or scope rationale.
Learning module content can be inaccurate. Please double-check its information. Report inaccuracies in PQC Today GitHub Discussions.