We Know Security but How Do We Secure GenAI End-to-End?
summary
Claims concrete fixes: sign models at train time, hash formats the same way used for signatures, and generate salsa provenance for each training cycle (data cleaning, fine‑tune, quantize) so supply‑chain tools can verify lineage. Demonstrates GitHub Actions proof‑of‑concept for model signing, cites SafeTensors replacing pickle to avoid RCE, and proposes using guac to merge model attestations, model cards and AI bomb manifests. Notes threats demonstrated (malware hidden in weights, prompt injection, build‑system compromise) and calls for ML package managers and provenance-aware model hubs.
AI-written summary · about
chapters (7)
- 0:00Introduction to GenAI security
- 1:45Common risks in AI and software
- 5:41ML supply chain vulnerabilities
- 8:50Current protection methods
- 12:25Model transparency standards
- 15:19Attacker tactics in ML
- 17:05Future security ideas
related talks
21mEnd-to-end secure ML Development
23mEnd-to-End Secure ML Development
19mModel transparency for AI/ML security
23mTrust, Track, and Verify: Securing AI Pipelines End-to-End
18mCommit Secure AI: Embedding Security in Code, Pipelines, and Production
35mGenAI Application Security: Not Just Prompt Injection
17mPreserving integrity and trust in AI
45mSecuring Software Supply Chain: Problems, Solutions, and AI/ML Challenges