--- schema: 1 type: video title: >- Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data - Sachin Kumar, LexisNexis description: >- You fine-tune LLMs and ship them. Your evals are green, your behavioral monitors are green — and a sleeper-agent backdoor can still flip the model to harmful output on a trigger you never tested. Behavioral testing can't reach it, and the interpretability tool people reach for — joint cross-model features (crosscoders) — dilutes the signal until it sits at the noise floor. The fix is in what the training data changed. A backdoor is a directional shift that fine-tuning writes into the model's activations, so you isolate it by watching the difference between the base and fine-tuned model. In a controlled SQL-injection backdoor, a sparse autoencoder trained on that difference flags it with 40× the signal of joint features, perfect precision, and zero false positives — from a single cheap layer. You'll leave knowing how to wire a "delta monitor" into your fine-tuning pipeline as a quiet CI gate. Based on my peer-reviewed paper accepted at IJCNN. Speakers: - Sachin Kumar (LexisNexis): Sachin Kumar is a Senior Data Scientist III and Tech Lead at LexisNexis, building agentic AI for the legal domain. His independent AI-safety and interpretability research has been accepted at top-tier venues including ACL, AAAI, and IJCNN. LinkedIn: https://www.linkedin.com/in/techsachinkumar/ GitHub: https://github.com/techsachinkr canonical: >- https://liquidgold.cc/channel/aidotengineer/explore/video/your-llm-deception-monitor-is-broken-the-fix-is-in-eeb61394c9a1 markdown: >- https://liquidgold.cc/channel/aidotengineer/video/your-llm-deception-monitor-is-broken-the-fix-is-in-eeb61394c9a1.md llms: null parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T08:32:58.779Z' language: null children: [] videoId: IQkVMvXQKLY durationSeconds: 838 publishedAt: '2026-07-08T08:05:21.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 11 endSeconds: 99 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=11' heldReason: rights - ordinal: 1 text: null startSeconds: 109 endSeconds: 199 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=109' heldReason: rights - ordinal: 2 text: null startSeconds: 109 endSeconds: 199 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=109' heldReason: rights - ordinal: 3 text: null startSeconds: 211 endSeconds: 299 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=211' heldReason: rights - ordinal: 4 text: null startSeconds: 211 endSeconds: 299 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=211' heldReason: rights - ordinal: 5 text: null startSeconds: 211 endSeconds: 299 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=211' heldReason: rights - ordinal: 6 text: null startSeconds: 211 endSeconds: 299 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=211' heldReason: rights - ordinal: 7 text: null startSeconds: 309 endSeconds: 444 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=309' heldReason: rights - ordinal: 8 text: null startSeconds: 309 endSeconds: 444 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=309' heldReason: rights - ordinal: 9 text: null startSeconds: 454 endSeconds: 534 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=454' heldReason: rights - ordinal: 10 text: null startSeconds: 534 endSeconds: 545 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=534' heldReason: rights - ordinal: 11 text: null startSeconds: 545 endSeconds: 671 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=545' heldReason: rights - ordinal: 12 text: null startSeconds: 545 endSeconds: 671 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=545' heldReason: rights - ordinal: 13 text: null startSeconds: 682 endSeconds: 796 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=682' heldReason: rights - ordinal: 14 text: null startSeconds: 682 endSeconds: 796 watch: 'https://liquidgold.cc/watch/IQkVMvXQKLY?t=682' heldReason: rights --- # Your LLM Deception Monitor Is Broken. The Fix Is in the Training Data - Sachin Kumar, LexisNexis You fine-tune LLMs and ship them. Your evals are green, your behavioral monitors are green — and a sleeper-agent backdoor can still flip the model to harmful output on a trigger you never tested. Behavioral testing can't reach it, and the interpretability tool people reach for — joint cross-model features (crosscoders) — dilutes the signal until it sits at the noise floor. The fix is in what the training data changed. A backdoor is a directional shift that fine-tuning writes into the model's activations, so you isolate it by watching the difference between the base and fine-tuned model. In a controlled SQL-injection backdoor, a sparse autoencoder trained on that difference flags it with 40× the signal of joint features, perfect precision, and zero false positives — from a single cheap layer. You'll leave knowing how to wire a "delta monitor" into your fine-tuning pipeline as a quiet CI gate. Based on my peer-reviewed paper accepted at IJCNN. Speakers: - Sachin Kumar (LexisNexis): Sachin Kumar is a Senior Data Scientist III and Tech Lead at LexisNexis, building agentic AI for the legal domain. His independent AI-safety and interpretability research has been accepted at top-tier venues including ACL, AAAI, and IJCNN. LinkedIn: https://www.linkedin.com/in/techsachinkumar/ GitHub: https://github.com/techsachinkr Canonical: https://liquidgold.cc/channel/aidotengineer/explore/video/your-llm-deception-monitor-is-broken-the-fix-is-in-eeb61394c9a1 · 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 (0:00 to 1:39) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 1: Document excerpt (1:39 to 3:19) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 2: Document excerpt (3:19 to 4:59) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 3: Document excerpt (4:59 to 7:24) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 4: Document excerpt (7:24 to 9:05) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 5: Document excerpt (9:05 to 11:11) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 6: Document excerpt (11:11 to 13:16) > Held: transcript rights do not permit this chunk's text. --- #### Chunk 7: Document excerpt (13:16 to 13:58) > Held: transcript rights do not permit this chunk's text.