Amir Feizpour
Michael Argast
Free Lesson

Open LLMs: Real Risks, Real Fixes

45 min
Oct 2, 2026 12:00 PM

By continuing, you agree to Maven's Terms and Privacy Policy.

What you'll learn

Map the real attack surface of open models

Distinguish exfiltration risk, weight tampering, and fine-tuning-based jailbreaks — they need different defenses.

Know why safety fine-tuning is fragile

Alignment can be stripped with ~10 adversarial examples for under $0.20 — plan defenses accordingly.

Choose the right deployment model for your risk profile

Compare closed-remote, open-remote, and open-local hosting on exfiltration, control, and operational burden.

Apply concrete mitigations today

Isolate system prompts, sign and pin supply-chain artifacts, and layer controls outside the model itself.

Why this topic matters

Open-weight models are spreading fast into production, but their security model differs sharply from closed APIs — fine-tuning can strip safety guardrails for pennies, and local deployment shifts the full security burden onto the adopting team. Builders and security leads need a practical framework for evaluating these tradeoffs before they ship, not after an incident.

You'll learn from

Amir Feizpour

Amir Feizpour

Founder @ Aggregate Intellect

Michael Argast

Michael Argast

CEO @ Kobalt.io

See all products from aggregate
Get free access