puodziukas.dev · remote · available Q3 2026

AI Engineer

FIPS 203/204/205. Adversarial ML. 0 false negatives. Post-quantum migration readiness.

FIPS 203/204/205 complete before NIST standards. Zero false negatives. Adversarial proof, not benchmark.

reproducible · verified · available Q3 2026

The problem

A model can pass every safety eval on the books and still fall to a red team in the first session. The gap between an eval suite and adversarial reality is exactly the gap that matters.

What exists, reproducible

The adversarial eval suite runs a fixed deterministic corpus: jailbreak, injection, hallucination, long-context coherence, data exfiltration. Zero false negatives across the public adversarial corpus (140 adversarial + 80 benign = 220, 140/140 caught, 0/80 over-blocked) means the suite is calibrated against an adversarial corpus, not ad-hoc tests.

Relevant projects

• →Adversarial eval suite on a fixed corpus: OWASP LLM Top 10 + custom injection vectors
• →FIPS 203/204/205 post-quantum migration readiness: standards and the 2030 NIST migration timeline mapped, harvest-now-decrypt-later exposure assessed
• →COBOL PICTURE-clause truncation detected: arithmetic overflow, years undetected by standard monitoring
• →Multi-agent council with SHA-256 hash-chain attestation: formal verification layer over LLM outputs
• →Model risk baseline harness: LLM outputs tied to a deterministic, replayable audit trail
• →Post-quantum key encapsulation: ML-KEM (FIPS 203), ML-DSA (FIPS 204), SLH-DSA (FIPS 205), all three
• →Harvest-now-decrypt-later threat model: board-level briefing artifact with risk quantification

What you receive in 30 days

1. Adversarial eval suite deployed: hallucination, injection, jailbreak, consistency (Week 1)
2. Model risk baseline: deterministic replay harness for regulated-deployment-adjacent scope
3. PQC readiness report: FIPS 203/204/205 gap analysis vs current stack

Companies at this intersection

Anthropic · OpenAI · Cohere · xAI · Imbue · Mistral AI · Together AI · Groq

All results under NDA. Representative figures. SHA-256 hash-chained audit trail available on request. · puodziukas.dev