Zyberpol — agentic security operations
An agentic SOC that investigates across every console and executes human-approved fixes from one screen — moving teams from monitoring to mitigation. Identity-first, human-approved, proof-native.
Co-Founder & Head of AI/ML · Vezran
Applied AI/ML, from detection to proof.
I build production ML and agentic security systems that don't stop at spotting the threat — they close it, with a human in command and evidence you can hand a regulator. Twelve years shipping real systems; now building the SOC that acts, not just alerts.
Currently: building Zyberpol — an agentic SOC that reasons, acts, and proves.
12+
years shipping
production ML
production ML
44+
merged pull requests
to core OSS libraries
to core OSS libraries
1
agentic SOC platform
in production
in production
3
companies —
2 senior roles, 1 founded
2 senior roles, 1 founded
Building
The part everyone skips: closing the threat, and proving it.
Most AI security stops at "we found something." At Vezran I lead the AI/ML behind Zyberpol — an agentic security-operations layer that takes an incident all the way to resolved, with a human approving the consequential moves and a tamper-proof record at the end.
ZYBERPOL · AGENTIC SOC
Detection got cheap. Trusted closure didn't.
Zyberpol sits on top of the security tools a team already runs — identity, endpoint, cloud, SIEM — correlates the noise into a handful of real incidents, recommends the safest containment, routes consequential actions through a human, then verifies the fix and exports proof an auditor accepts. The model is rented and swappable; the control layer and its governance are the moat.
01
Detect
Correlate alert noise into a few high-priority incidents.
02
Decide
Auto-gather evidence; recommend the safest containment.
03
Approve
Route consequential actions through a human in command.
04
Act & verify
Execute, then confirm the fix actually worked.
05
Prove
Export a tamper-proof record for audits and insurers.
Selected work
Systems shipped, and risks found.
From agentic security platforms to production ML at enterprise scale, to coordinated vulnerability disclosures in the tools everyone builds on.
Adversarial AI & disclosure
Coordinated disclosures in HuggingFace Transformers, and benchmarks probing prompt-injection robustness with DSPy and AgentDojo.
Contact-center NLP at scale
Production natural-language systems for a global contact-center operation — routing, understanding, and surfacing signal from millions of customer interactions.
Package-risk modeling
Risk models across a logistics network at national scale — predicting and flagging shipment risk to protect operations and cost.
Open source
I don't just use the tools — I fix them.
44 merged · 14 open
accepted contributions across 11 external ML & AI-security projects
Writing & notes
Field notes on applied AI.
Long-form on the problems that actually bite in production — and short notes from the bench.
My fraud detector scored 100% in every language. It wasn't reading any of them.
Defanging the dataset handed the model a shortcut that made it look multilingual. Strip the placeholder and non-Latin scripts collapse — Chinese 1.00 → 0.09, Arabic 0.98 → 0.04. The failure wasn't in the model. It was in the measurement.
Read the essay →Upgrading your AI model can silently break its security
A more capable model started following attackers' orders. Across eight models, nothing — size, vendor, or capability — predicted which would break.
Read →I built a benchmark for AI-written fraud — and the first result was too good to be true
LureBench. Control for the dataset confound and telling AI-written fraud from human is close to a coin flip. The defensive problem is still open.
Read →When does prompt optimization stop helping and start hurting security?
Three regimes by base-model strength, with a practitioner decision rule — DSPy optimizers vs AgentDojo prompt-injection attacks.
Read →The DP-LoRA silent corruption: 5 months of broken fine-tuning, hidden in plain sight
A device-placement ordering quirk between opacus, PEFT, and HuggingFace zeroed gradients while noise kept accumulating. What to check in your own setup.
Read →Does DSPy prompt optimization weaken adversarial robustness?
Optimizing prompts for accuracy tightens them to the training distribution — and can cost ~20 points of security on the harder attacks.
Read →
Contact
Building something at the edge of
Building something at the edge of
AI and security?
I'm always up for a sharp conversation — agentic systems, LLM security, or turning a security team from monitoring into mitigation.