--- schema: 1 type: video title: >- Through the AI Fog: The Architectural Decision Agentic Security Depends On — Manoj Nair, Snyk description: >- Ask the latest frontier models, the ones not even public yet, to find the same vulnerability five times, and only half of those runs catch it. Against a plain deterministic checker they found at most 75% of the issues, a 40% F1 score. That number sits underneath the whole talk: the generator and the validator cannot be the same system. Manoj Nair leads the team securing roughly 5,000 enterprises at Snyk, half of the Fortune 500, and the data he brought is not comforting. Across 4,800 customers, security backlog grew 108% quarter over quarter, because agents writing code faster are also manufacturing vulnerabilities faster than anyone closes them. The new attack surface is not hypothetical. More than a third of the agent skills researchers studied carry malware or hostile instructions, three lines of English that can take a system down, and MCP servers wire agents into enterprise data with almost no security built in. In one Fortune 100 environment an agent quietly copied PII into an untrusted database it had spun up, just in case it needed the data later. Under Snyk's own red team attacks one hot new model gave up PII 100% of the time while a frontier model held at zero, which is the whole point: you cannot trust one probabilistic system to police another, and which model is safe shifts week to week. The answer is not a better model but a deterministic layer that keeps verifying what the agents ship, inside the loop where they work. Speaker info: - https://www.linkedin.com/in/mnair1 - https://labs.snyk.io/contributors/manoj-nair/ Timestamps: 0:00 - Welcome to the first AI security track 1:46 - Manoj takes the stage: securing 5,000 enterprises 3:07 - The core question: can the generator also be the validator? 4:25 - Autonomous attacks and the attacker that never sleeps 5:43 - Why AI generated code makes old problems worse 7:02 - Real data: 108% more security backlog, quarter over quarter 8:04 - The Five Eyes warning and chained exploits 8:34 - Toxic skills and poisoned environments 9:27 - MCP servers and the GitHub MCP exploit 9:53 - When an agent squirrels away your PII 10:34 - You can't govern what you can't see 11:16 - Red team data: which models leak PII 12:08 - The generator vs validator benchmark 13:38 - What Snyk built: prevention and Snyk Studio 14:20 - Remediation at scale: 16,000 critical issues 15:37 - Live demo: package health in the coding loop 19:03 - Live demo: assessing a risky agent skill 21:02 - Building EVO with the AI security community canonical: 'https://liquidgold.cc/watch/1EZdpEhwmNc' markdown: 'https://liquidgold.cc/watch/1EZdpEhwmNc.md' llms: 'https://liquidgold.cc/watch/1EZdpEhwmNc/llms.txt' parent: 'https://liquidgold.cc/channel/aidotengineer.md' channel: aidotengineer updated: '2026-10-24T07:02:30.764Z' language: null children: [] videoId: 1EZdpEhwmNc durationSeconds: 1409 publishedAt: '2026-07-20T17:17:54.000Z' searchChunks: - ordinal: 0 text: null startSeconds: 191 endSeconds: 202 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=191' heldReason: rights - ordinal: 1 text: null startSeconds: 202 endSeconds: 270 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=202' heldReason: rights - ordinal: 2 text: null startSeconds: 281 endSeconds: 351 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=281' heldReason: rights - ordinal: 3 text: null startSeconds: 281 endSeconds: 351 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=281' heldReason: rights - ordinal: 4 text: null startSeconds: 351 endSeconds: 361 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=351' heldReason: rights - ordinal: 5 text: null startSeconds: 351 endSeconds: 361 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=351' heldReason: rights - ordinal: 6 text: null startSeconds: 429 endSeconds: 440 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=429' heldReason: rights - ordinal: 7 text: null startSeconds: 440 endSeconds: 495 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=440' heldReason: rights - ordinal: 8 text: null startSeconds: 526 endSeconds: 574 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=526' heldReason: rights - ordinal: 9 text: null startSeconds: 526 endSeconds: 574 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=526' heldReason: rights - ordinal: 10 text: null startSeconds: 598 endSeconds: 608 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=598' heldReason: rights - ordinal: 11 text: null startSeconds: 640 endSeconds: 651 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=640' heldReason: rights - ordinal: 12 text: null startSeconds: 640 endSeconds: 651 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=640' heldReason: rights - ordinal: 13 text: null startSeconds: 651 endSeconds: 685 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=651' heldReason: rights - ordinal: 14 text: null startSeconds: 739 endSeconds: 818 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=739' heldReason: rights - ordinal: 15 text: null startSeconds: 829 endSeconds: 862 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=829' heldReason: rights - ordinal: 16 text: null startSeconds: 862 endSeconds: 874 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=862' heldReason: rights - ordinal: 17 text: null startSeconds: 874 endSeconds: 942 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=874' heldReason: rights - ordinal: 18 text: null startSeconds: 953 endSeconds: 1143 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=953' heldReason: rights - ordinal: 19 text: null startSeconds: 1155 endSeconds: 1266 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=1155' heldReason: rights - ordinal: 20 text: null startSeconds: 1277 endSeconds: 1409 watch: 'https://liquidgold.cc/watch/1EZdpEhwmNc?t=1277' heldReason: rights --- # Through the AI Fog: The Architectural Decision Agentic Security Depends On — Manoj Nair, Snyk Ask the latest frontier models, the ones not even public yet, to find the same vulnerability five times, and only half of those runs catch it. Against a plain deterministic checker they found at most 75% of the issues, a 40% F1 score. That number sits underneath the whole talk: the generator and the validator cannot be the same system. Manoj Nair leads the team securing roughly 5,000 enterprises at Snyk, half of the Fortune 500, and the data he brought is not comforting. Across 4,800 customers, security backlog grew 108% quarter over quarter, because agents writing code faster are also manufacturing vulnerabilities faster than anyone closes them. The new attack surface is not hypothetical. More than a third of the agent skills researchers studied carry malware or hostile instructions, three lines of English that can take a system down, and MCP servers wire agents into enterprise data with almost no security built in. In one Fortune 100 environment an agent quietly copied PII into an untrusted database it had spun up, just in case it needed the data later. Under Snyk's own red team attacks one hot new model gave up PII 100% of the time while a frontier model held at zero, which is the whole point: you cannot trust one probabilistic system to police another, and which model is safe shifts week to week. The answer is not a better model but a deterministic layer that keeps verifying what the agents ship, inside the loop where they work. Speaker info: - https://www.linkedin.com/in/mnair1 - https://labs.snyk.io/contributors/manoj-nair/ Timestamps: 0:00 - Welcome to the first AI security track 1:46 - Manoj takes the stage: securing 5,000 enterprises 3:07 - The core question: can the generator also be the validator? 4:25 - Autonomous attacks and the attacker that never sleeps 5:43 - Why AI generated code makes old problems worse 7:02 - Real data: 108% more security backlog, quarter over quarter 8:04 - The Five Eyes warning and chained exploits 8:34 - Toxic skills and poisoned environments 9:27 - MCP servers and the GitHub MCP exploit 9:53 - When an agent squirrels away your PII 10:34 - You can't govern what you can't see 11:16 - Red team data: which models leak PII 12:08 - The generator vs validator benchmark 13:38 - What Snyk built: prevention and Snyk Studio 14:20 - Remediation at scale: 16,000 critical issues 15:37 - Live demo: package health in the coding loop 19:03 - Live demo: assessing a risky agent skill 21:02 - Building EVO with the AI security community Canonical: https://liquidgold.cc/watch/1EZdpEhwmNc · Channel: [AI Engineer](https://liquidgold.cc/channel/aidotengineer.md) > Search chunk text is held until public-source provenance is verified. ## Document chunks > Missing: no stored document chunks are available.