--- schema: 1 type: video title: >- $1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs – Diego Carpintero description: >- LLM-based attacks are no longer the exception, they are the baseline. This talk maps the six most common attack vectors found in production AI systems: Prompt and Context Injection, Model Internals, RAG Poisoning, MCP Exploits, and Agentic Escalation. From there, it dives into the architecture of ModernBERT and shows how to fine-tune it into a lightweight, self-hosted guardrails layer for under a dollar. What you will learn: - The Zero Trust Gap in LLMs: what these attack vectors share in common, and why model alignment and human review alone are not enough - The secret sauce that makes encoder models beat LLM-as-a-Judge in latency and flexibility - ModernBERT under the hood: a deep dive into Alternating Attention, Unpadding & Sequence Packing, RoPE, and FlashAttention - Building your own safety layer: a practical walkthrough of fine-tuning ModernBERT as a safety discriminator - Live Demo: real attack prompts from each vector tested against our model" Speaker info: - Diego Carpintero - AI Engineer | Tech Entrepreneur | Open Source Contributor | NVIDIA Certified Professional (NCP-GENL) Timestamps: 00:00 Intro --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/YZHPEkfy2kc-1-ai-guardrails-the-unreasonable-effectiveness-of-finetuned-modernberts-diego-carpentero Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/1-ai-guardrails-the-unreasonable-effectiveness-of-f-14e10444d1d6 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/1-ai-guardrails-the-unreasonable-effectiveness-of-f-14e10444d1d6.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T15:26:19.481Z' language: null children: [] videoId: YZHPEkfy2kc durationSeconds: 2633 publishedAt: '2026-04-16T11:00:07.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 1 endSeconds: 12 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1' heldReason: rights - ordinal: 1 text: null startSeconds: 12 endSeconds: 115 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=12' heldReason: rights - ordinal: 2 text: null startSeconds: 12 endSeconds: 115 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=12' heldReason: rights - ordinal: 3 text: null startSeconds: 125 endSeconds: 261 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=125' heldReason: rights - ordinal: 4 text: null startSeconds: 125 endSeconds: 261 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=125' heldReason: rights - ordinal: 5 text: null startSeconds: 125 endSeconds: 261 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=125' heldReason: rights - ordinal: 6 text: null startSeconds: 272 endSeconds: 395 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=272' heldReason: rights - ordinal: 7 text: null startSeconds: 272 endSeconds: 395 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=272' heldReason: rights - ordinal: 8 text: null startSeconds: 272 endSeconds: 395 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=272' heldReason: rights - ordinal: 9 text: null startSeconds: 406 endSeconds: 485 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=406' heldReason: rights - ordinal: 10 text: null startSeconds: 495 endSeconds: 575 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=495' heldReason: rights - ordinal: 11 text: null startSeconds: 587 endSeconds: 689 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=587' heldReason: rights - ordinal: 12 text: null startSeconds: 587 endSeconds: 689 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=587' heldReason: rights - ordinal: 13 text: null startSeconds: 699 endSeconds: 839 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=699' heldReason: rights - ordinal: 14 text: null startSeconds: 839 endSeconds: 849 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=839' heldReason: rights - ordinal: 15 text: null startSeconds: 849 endSeconds: 967 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=849' heldReason: rights - ordinal: 16 text: null startSeconds: 849 endSeconds: 967 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=849' heldReason: rights - ordinal: 17 text: null startSeconds: 978 endSeconds: 1066 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=978' heldReason: rights - ordinal: 18 text: null startSeconds: 978 endSeconds: 1066 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=978' heldReason: rights - ordinal: 19 text: null startSeconds: 1078 endSeconds: 1168 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1078' heldReason: rights - ordinal: 20 text: null startSeconds: 1078 endSeconds: 1168 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1078' heldReason: rights - ordinal: 21 text: null startSeconds: 1178 endSeconds: 1267 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1178' heldReason: rights - ordinal: 22 text: null startSeconds: 1267 endSeconds: 1278 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1267' heldReason: rights - ordinal: 23 text: null startSeconds: 1358 endSeconds: 1370 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1358' heldReason: rights - ordinal: 24 text: null startSeconds: 1496 endSeconds: 1507 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1496' heldReason: rights - ordinal: 25 text: null startSeconds: 1605 endSeconds: 1684 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1605' heldReason: rights - ordinal: 26 text: null startSeconds: 1694 endSeconds: 1794 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1694' heldReason: rights - ordinal: 27 text: null startSeconds: 1909 endSeconds: 1999 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=1909' heldReason: rights - ordinal: 28 text: null startSeconds: 2009 endSeconds: 2115 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2009' heldReason: rights - ordinal: 29 text: null startSeconds: 2115 endSeconds: 2127 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2115' heldReason: rights - ordinal: 30 text: null startSeconds: 2115 endSeconds: 2127 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2115' heldReason: rights - ordinal: 31 text: null startSeconds: 2208 endSeconds: 2218 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2208' heldReason: rights - ordinal: 32 text: null startSeconds: 2335 endSeconds: 2416 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2335' heldReason: rights - ordinal: 33 text: null startSeconds: 2416 endSeconds: 2426 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2416' heldReason: rights - ordinal: 34 text: null startSeconds: 2426 endSeconds: 2552 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2426' heldReason: rights - ordinal: 35 text: null startSeconds: 2562 endSeconds: 2633 watch: 'https://liquidgold.cc/watch/YZHPEkfy2kc?t=2562' heldReason: rights --- # $1 AI Guardrails: The Unreasonable Effectiveness of Finetuned ModernBERTs – Diego Carpintero LLM-based attacks are no longer the exception, they are the baseline. This talk maps the six most common attack vectors found in production AI systems: Prompt and Context Injection, Model Internals, RAG Poisoning, MCP Exploits, and Agentic Escalation. From there, it dives into the architecture of ModernBERT and shows how to fine-tune it into a lightweight, self-hosted guardrails layer for under a dollar. What you will learn: - The Zero Trust Gap in LLMs: what these attack vectors share in common, and why model alignment and human review alone are not enough - The secret sauce that makes encoder models beat LLM-as-a-Judge in latency and flexibility - ModernBERT under the hood: a deep dive into Alternating Attention, Unpadding & Sequence Packing, RoPE, and FlashAttention - Building your own safety layer: a practical walkthrough of fine-tuning ModernBERT as a safety discriminator - Live Demo: real attack prompts from each vector tested against our model" Speaker info: - Diego Carpintero - AI Engineer | Tech Entrepreneur | Open Source Contributor | NVIDIA Certified Professional (NCP-GENL) Timestamps: 00:00 Intro --- Watch along with the corrected transcript, summary, timestamps and resources on this talk's official AIE page: https://ai.engineer/talks/YZHPEkfy2kc-1-ai-guardrails-the-unreasonable-effectiveness-of-finetuned-modernberts-diego-carpentero Join our upcoming AI Engineer conferences (see https://ai.engineer landing page): - AIEi Shanghai, November 5–6 — the first for meeting top Chinese labs and engineers - AIE CODE in SF, November 10–12 — the top AI Coding conference returns - AIEi Sydney, December 7–8 — the first AIEi run alongside NeurIPS 2026 in Sydney --- Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/1-ai-guardrails-the-unreasonable-effectiveness-of-f-14e10444d1d6 · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks --- #### Chunk 0: Document excerpt (0:01 to 0:12) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (0:12 to 1:55) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (1:55 to 2:05) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (2:05 to 4:21) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (4:21 to 4:32) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (4:32 to 6:35) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (6:35 to 8:05) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (8:05 to 9:35) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 8: Document excerpt (9:35 to 9:47) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 9: Document excerpt (9:47 to 11:29) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 10: Document excerpt (11:29 to 11:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 11: Document excerpt (11:39 to 13:59) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 12: Document excerpt (13:59 to 14:09) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 13: Document excerpt (14:09 to 16:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 14: Document excerpt (16:07 to 17:46) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 15: Document excerpt (17:46 to 19:28) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 16: Document excerpt (19:28 to 21:07) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 17: Document excerpt (21:07 to 22:38) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 18: Document excerpt (22:38 to 22:50) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 19: Document excerpt (22:50 to 24:56) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 20: Document excerpt (24:56 to 26:34) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 21: Document excerpt (26:34 to 28:04) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 22: Document excerpt (28:04 to 28:14) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 23: Document excerpt (28:14 to 29:54) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 24: Document excerpt (29:54 to 31:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 25: Document excerpt (31:39 to 33:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 26: Document excerpt (33:19 to 35:15) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 27: Document excerpt (35:15 to 36:48) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 28: Document excerpt (36:48 to 38:43) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 29: Document excerpt (38:43 to 40:26) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 30: Document excerpt (40:26 to 42:32) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 31: Document excerpt (42:32 to 43:53) > Held: transcript rights do not permit this chunk's text.