Hard problems
RNA folding as a security assumption.
Predicting the three-dimensional structure of an RNA molecule from its sequence is an extremely difficult problem, including for a quantum computer.
Paris, France
AI systems for
cryptographic
security.
We are developing new approaches to post-quantum cryptanalysis and offensive security using evolutionary program synthesis. Our goal is to build systems that discover exploitable structures in cryptographic schemes, and in the software that implements them, before adversaries do.
01Research
Post-quantum cryptography is being standardized. Whether the standardized schemes are secure against AI-augmented attacks is an open question.
Agades studies post-quantum cryptanalysis as a search problem: how can AI generate, test, and improve attack strategies against lattice-based cryptographic schemes before those weaknesses become operational risks?
-
Standards
2024
NIST finalized its first post-quantum standards. ML-KEM and ML-DSA rest on lattice problems: Learning With Errors and the Shortest Vector Problem.
-
Migration
2026 → 2030
The EU roadmap has every Member State starting its transition by the end of 2026, and critical infrastructure migrated by 2030.
-
ML cryptanalysis
2022 →
SALSA, PICANTE, VERDE and FRESCA recover LWE secrets with machine learning. Still far from production parameters, but the tools are getting better.
02AgadesEvolve
AgadesEvolve
A language model proposes programs, an objective evaluator scores them, the best survive and recombine.
An attack on a post-quantum scheme is a program: a combination of algebraic techniques, lattice reduction strategies and parameter choices that, composed in the right order, lowers the security of the scheme below its claimed level.
The evaluator is the Lattice Estimator, which returns the cost of a strategy in security bits. The objective is to minimize that number. If an evolved strategy reaches 2115 where known strategies reach 2128, that is a result.
- Status
- Developed
- Models
- Any frontier model, closed or open-weight
- Oracle
- Lattice Estimator, cost in security bits
02Architecture
The optimization loop.
Each iteration builds a context from an archive of elites, asks a frontier LLM for a new candidate, and scores it in three stages. A candidate that beats the elite of its cell takes its place. Search escalates from local exploitation to bandit-driven exploration, then to injected literature, when the fitness stagnates.
N1 local exploitation
N2 exploration (UCB bandit)
N3 injected literature
routing · escalation if F stagnates
Archive A : b(x) → elite
Context c
parents +
inspirations +
literature
inspirations +
literature
Frontier LLM
x′ ~ π( · | c )
x′ ~ π( · | c )
1 Attack-plan validity
2 f(x′) : estimated attack cost
3 Replay on reduced instances
Lineage Inspired by AlphaEvolve (Google DeepMind, 2025): language models as mutation operators, programs scored against a fitness oracle.
03Offensive security
The same search architecture applies to vulnerability discovery in software and firmware.
The same architecture can search software and firmware for exploitation paths that static signatures miss, using program structure and real target behavior as the scoring signal. In cryptanalysis, the oracle is the Lattice Estimator. In software, the oracle is the target system itself.
SecAgent: from binary to migration plan.
SecAgent is the program-analysis layer. Its first pipeline takes firmware and binaries to a cryptographic bill of materials and a post-quantum migration plan. Offline, without an LLM, replayable.
- Inputsfirmware, binaries, protocols
- Ingestionbyte lineage, fingerprints
- Analysisbinary IR, CFG, identified crypto
- CBOMCycloneDX, with evidence
- Exposurecode, protocols, equipment
- Migrationsigned debt, exact Pareto plan
Every finding carries one of four statuses. What is not proven stays unknown.
- Confirmed
- Refuted under model
- Unknown
- Heuristic
04Readiness
What exists, what remains to be proven.
| System | Designed | Developed | Tested internally | Validated externally | Next proof |
|---|---|---|---|---|---|
| AgadesEvolveevolutionary engine | Done | Current stage | Next step | Not yet | Large-scale campaigns |
| SecAgentpost-quantum migration | Done | Done | Current stage | Next step | Pilots on real industrial fleets |
| Agades PQC GymLWE/MLWE, open source | Done | Done | Current stage | Next step | Review by cryptographers |
| Lattice Estimator5 LWE attacks reproduced | Done | Done | Current stage | Next step | ML-KEM case study |
Next proof. Compute and case studies are the bottleneck, not the architecture.
05Beyond lattices
Biology as a source of cryptographic hardness.
Agades is also interested in biological complexity as a possible source of cryptographic hardness, from RNA folding to DNA-based physical primitives.
Read the notePhysical substrate
DNA as a key that cannot be copied.
Chemical unclonable functions use pools of modified DNA molecules for authentication and anti-counterfeiting, physically bound to an object.
06Notes
Notes.
Long-form writing on the research program. Case studies and technical write-ups will be published here.
- Research note · Cryptanalysis Is post-quantum cryptography secure against AI-augmented attacks? Read
- Research note · Offensive security Vulnerability discovery as a search over programs. Read
- Research note · Post-quantum migration Post-quantum migration for IoT fleets. Read
- Research note · Beyond lattices Biology as a source of cryptographic hardness. Read
07Team
Zakaria Laabsi
Founder · AI/ML Engineer
The cryptography that will protect data for the next decades is being standardized and deployed now, at the moment AI is changing how attacks are found. That is why I started Agades: to put AI to work on these schemes, and on the software that implements them, before adversaries do.
08Contact
Get in touch.
Investors, research collaborators and security teams: write to us directly.
Security team evaluating post-quantum exposure? See engagements →