Deterministic infrastructure safety

Find the blast radius
before it ships.

Two shifts are hitting production at once. In a post-quantum migration, ML-DSA signatures are 52× larger than the Ed25519 signatures they replace, and they will overflow limits nobody remembers setting. Meanwhile AI agents read untrusted context, and RAG poisoning spreads through your tickets, code and vector stores. Tokenkeep traces both, byte by byte and edge by edge, and gives you an exact, auditable answer.

Post-quantum migration AI agent rollback
  • Deterministic. Integer-exact verdicts, with no LLM in the decision path.
  • Auditable. Every finding points to a file and line, and every rollback to a signed edge.
PQC

Post-quantum migration analysis

Traces every signature, key and ciphertext from the code that produces it to the byte limits it must fit through: SQL columns, cookies, header buffers, TLS records, firmware stacks. Reports every overflow, then hands you the patch.

  • ML-DSA
  • SLH-DSA
  • FN-DSA
  • ML-KEM
  • Hybrids
Explore PQC → Epistemic

RAG poisoning rollback for AI agents

Attributes what your agents wrote to what they read and keeps that lineage as a signed DAG. When a source turns out to be poisoned, it severs retrieval and reverts the code it contaminated, in that order, with proof that nothing is left.

  • Vector stores
  • Tickets
  • Git
  • MCP
Explore Epistemic →

Free reference

Post-quantum sizes, worked out to the byte.

Signature, key and ciphertext sizes for ML-DSA, SLH-DSA, ML-KEM and FN-DSA, encoded the way your stack actually stores them, and checked against the limits they hit.

Writing

Technical note · 8 min read

ML-DSA certificate chains: the 184 bytes that break TLS handshakes

A three-certificate ML-DSA-65 chain comes to 16,568 bytes. mbedTLS's default record buffer holds 16,384. We work through the arithmetic byte by byte: why your CBOM won't flag it, and the four ways to fix it.

Read the note →

Approach

Security tools should be as exact as the systems they protect.

01

Exact, not estimated

Byte counts come from FIPS 203/204/205 tables and real encoder arithmetic. Attribution verdicts are integer math. The same input always gives the same output.

02

Read before write

Diagnostics are read-only by construction. Changes ship as reviewed diffs, and rollbacks verify each phase before starting the next.

03

Proven, then shipped

Property tests, fuzzing, model checks and closed-loop suites: apply the fix, re-analyze, and confirm zero findings. Written in Rust.

Start here

Find out what breaks before your users do.

Planning a PQC migration, or running agents against production context? Tell us your stack. We'll show you the blast radius.

or email