AI
Work Decomposed Into Executable Blocks
Miessler’s thesis: every job, including security work, is fragmenting into discrete AI-executable tasks he calls Lego blocks. The blocks connect and compose into workflows. Knowledge work in the US costs $10 trillion in annual compensation. Globally, $50 to $70 trillion. That is the addressable market driving AI investment. Breaking down work into transparent, text-describable tasks gives AI the structure it needs to execute reliably. Miessler spent 18 months writing about this pattern before Claude Code crystallized it into practice.
How Context Reduces the Expertise Required
A junior SOC analyst handed a bare alert struggles to answer “is this malicious?” The same analyst, walked up to a whiteboard showing a week’s worth of timeline, stolen credential chain, and connection logs, can answer quickly. Miessler’s framing: more context means less expertise required to do the work. He formalized this as Unified Entity Context, a data lake for decision-making. That same principle explains why Claude Code outperforms raw Claude models. The context management does the heavy lifting.
From Prompt Libraries to Agent Skills
Miessler’s 2023 Fabric project started this pattern. Fabric is 250 crowdsourced prompts, each doing one task well. It hit the top of GitHub and stayed there. Claude Code arrived in March 2025. Miessler initially dismissed it as a command-line wrapper. Then Anthropic shipped agent skills, markdown directories that bundle instructions, context, and code into a single reusable package. Hundreds appeared on GitHub within hours. Miessler migrated his entire stack within a few hours. Skills are designed to be invoked by agents, not just humans.
How Agentic AI Handles Security Program Reporting
Miessler demoed Kai, his personal AI platform, fielding 20 specialized agents against a fake authentication company’s data. The context store is plain markdown and CSV files, editable by anyone on the team. A question about program status returned a full metrics breakdown in two and a half minutes. A follow-up request to build a web dashboard took three minutes and produced working HTML. When corporate changes a goal in one file, the entire system reflects that change on the next query.
Five Levels of AI Adoption
Miessler laid out a five-level maturity model for AI in business. Level 0 is pre-ChatGPT. Level 2 is chatbots with manual follow-through. Level 3 is agentic platforms like Kai, which he calls the last heavily human-focused stage. Level 4 turns those agentic queries into persistent, continuously running pipelines. He put widespread level-4 adoption at around 2027. Level 5 is the AI managing the gap between current and desired state on its own, continuously. Some attackers are already at level 2 through 4.
The Attacker Is on the Same Path
Defenders are not the only ones building Lego-block workflows. Anthropic tracked one campaign that looked like a 10-person team operating for months. One attacker ran it in a week using AI automations. The gap between defender and attacker AI maturity is the real risk surface. Miessler’s advice: reach level 3 quickly, build a personal agent platform, and capture your own hard-won expertise as skills before a generic version shows up on GitHub and commoditizes it.
Q&A
When do we set up AI controls? Build and secure simultaneously; there is no safe-first phase. ▶ 32:34
How do you differentiate yourself when AI commoditizes expertise? Encode your hard-earned, experience-shaped taste into private skills that others cannot replicate. ▶ 34:01
How do you prevent hallucinations from corrupting the system? Use a second AI with different goals to check the first; some human oversight remains. ▶ 36:16
What evidence shows attackers are already further along? Anthropic observed one attacker who, using AI automations, did in one week what looked like a 10-person team operating for months. ▶ 43:57
Notable Quotes
I mean, they just came out and said it. Daniel Miessler · ▶ 12:09
highly awesome and highly scary. Daniel Miessler · ▶ 28:10
they did it in a week. So they had Daniel Miessler · ▶ 44:24
Key Takeaways
- All security work is fragmenting into AI-executable task blocks, similar to Lego pieces connecting.
- More context in AI systems means less expertise needed to answer the same security question.
- Attackers using AI at level 3 or 4 multiply their operational reach by orders of magnitude.
About the Speaker(s)
Daniel Miessler is a cybersecurity and AI engineer and founder based in the San Francisco Bay Area. He has written online since 1999 and created SecLists, the open-source payload repository used by penetration testers and bug bounty hunters worldwide. He leads the OWASP Internet of Things project and has spoken at DEF CON, Black Hat, and SANS. His stated goal is to help build what he calls Human 3.0, a positive, human-based AI future.