![Dan Guido - 200 Bugs/Week/Engineer: How We Rebuilt Trail of Bits Around AI | [un]prompted 2026](https://wm-img.halpindev.com/p-briefread_c-10_b-10/urlb/aHR0cDovL2ltZy55b3V0dWJlLmNvbS92aS9rZ3d2QXlGN3FzQS9tYXhyZXNkZWZhdWx0LmpwZw==.jpg)
TLDR;
Dan Guido, CEO of Trail of Bits, discusses the company's journey to becoming AI-native. He addresses the common issue of companies failing to see measurable impact from AI due to simply adding tools without changing the underlying system. He defines three levels of AI integration: AI-assisted, AI-augmented, and AI-native, emphasizing the importance of designing an organization from the ground up with AI as a co-participant. Guido also shares the psychological barriers encountered during this transformation, such as self-enhancing bias, identity threat, opacity, and intolerance for imperfection, and the strategies used to overcome them. He details the five-part system Trail of Bits built, including standardization, an AI handbook, a maturity matrix, hackathons, and skills repositories, to foster AI adoption and improve productivity.
- AI native is an operating system, not a tool
- Psychological barriers to AI adoption: self-enhancing bias, identity threat, opacity, and intolerance for imperfection
- Five-part system: standardization, AI handbook, maturity matrix, hackathons, and skills repositories
Introduction [0:05]
Gadi introduces Dan Guido, the CEO and co-founder of Trail of Bits, highlighting Guido's pioneering work in making Trail of Bits an AI-native company. Gadi emphasizes that Trail of Bits has not only embraced AI but has also invested heavily in open-source contributions and AI security research, positioning themselves as leaders in the field. He expresses his admiration for Guido's vision and the company's transformation, noting that Guido's approach combines technical expertise with a thoughtful ethos.
AI-Assisted vs. AI-Native [1:55]
Dan Guido begins by referencing a recent survey indicating that AI has had no measurable impact on employment or productivity for many CEOs, a phenomenon economists are calling the "Solo Paradox." He argues that this is because most companies are simply adding AI tools without fundamentally changing their systems. Guido differentiates between AI-assisted, where AI is merely a tool, and AI-native, where AI is an integral part of the operating system. The rest of the talk will focus on how Trail of Bits is building an AI-native operating system.
Defining AI-Native [3:16]
Guido defines three levels of AI integration. AI-assisted is the basic level where people use tools like ChatGPT to draft emails and generate boilerplate, without changing organizational structure or workflows. AI-augmented involves incrementally redesigning workflows by incorporating AI agents, such as having AI perform the first pass on code reviews. AI-native represents a structural shift where the organization is designed from the ground up with AI as a co-participant, integrating AI into knowledge management, delivery models, and expertise. At Trail of Bits, this means that their security expertise compounds with code, and every engineer uses specialized AI agents built from their extensive audit knowledge.
Overcoming Resistance to AI Adoption [4:43]
Guido candidly discusses the significant pushback he faced when launching the AI initiative at Trail of Bits, with only a small percentage of the company initially supporting the change. He notes that resistance to new technology is common, especially among senior experts. To address this, Guido focused on overcoming four key psychological barriers: self-enhancing bias, identity threat, opacity, and intolerance for imperfection. The core insight is that the problem is not the technology itself, but the people's perception and acceptance of it.
Psychological Barriers to AI Adoption [5:33]
Guido elaborates on the four psychological barriers to AI adoption. Self-enhancing bias is the tendency to perceive oneself favorably, leading individuals to take credit for successes and blame external factors for failures. Identity threat arises when people feel that AI undermines their expertise and experience. Opacity refers to the lack of understanding and trust in AI judgment compared to human judgment. Intolerance for imperfection means that people quickly abandon AI systems after a single error, even if the AI is generally more accurate than humans.
Strategies to Overcome Barriers [8:56]
To address self-enhancing bias, Trail of Bits created a maturity matrix that provides a visible ladder for improvement, making conversations about AI adoption more concrete and creating social proof as peers advance. To combat identity threat, they focused on skill-allowing automations rather than skill-replacing ones, framing AI as a way to enhance expertise. Hackathons were used instead of mandates to allow people to experiment with AI on their own terms. To mitigate intolerance for imperfection, they invested in reducing AI failures and provided ways for people to modify the algorithms. For opacity, they clearly articulated which tools were allowed and why, making adoption fast and visible to reinforce the company's commitment to AI.
The Five-Part System for AI Integration [12:00]
Guido introduces the five-part system Trail of Bits built to become AI-native, emphasizing that all five parts are necessary for compounding success. The first part is standardization, which involves getting everyone on Claude Code with supported configurations and clear guidelines. The second part is an AI handbook that removes ambiguity by specifying which tools are approved for sensitive data. The third part is an AI maturity matrix that makes AI a first-class professional capability with clear levels and expectations. The fourth part is hackathons, which serve as a management system for short, focused sprints. The fifth part is skills repositories, which capture reusable workflows and artifacts.
Standardization and AI Handbook [12:34]
Guido emphasizes the importance of standardization, noting that without it, companies end up with fragmented workflows and no leverage. Trail of Bits standardized on Claude Code, treating it like any other enterprise tool with supported configurations and clear guidelines. They also created an AI handbook to remove ambiguity about which tools are approved, especially for sensitive data. This handbook helps project managers and engineers maintain consistent client conversations and provides clear answers to questions about data usage.
AI Maturity Matrix and Hackathons [13:48]
The AI maturity matrix makes AI a formal professional capability with clear levels and expectations, integrating it into performance reviews and resource allocation. Hackathons are used as a management system for short, focused sprints, allowing the company to keep pace with the rapidly changing AI ecosystem. These hackathons include constraints, such as running Claude Code in bypass permissions mode, to encourage the use of sandboxing and guardrails. Success is measured by activity, such as issues fixed and PRs reviewed, and people work in pairs to ensure quality control.
Skills Repositories and Curated Marketplace [16:21]
Skills repositories capture the motion created by hackathons, storing reusable workflows with examples, constraints, and verification methods. Trail of Bits maintains three skills repos: an internal one for company-specific workflows, an external one for community validation, and a curated marketplace. The curated marketplace provides a safe supply chain for third-party skills, ensuring that employees use verified and approved plugins. This helps prevent the use of random and potentially harmful AI tools.
Simplified Defaults, Sandboxing, and Package Cooldown [18:07]
To make AI adoption easier, Trail of Bits created dead-simple defaults, publishing a repository with recommended Claude Code configurations and usage instructions. They also provide a variety of sandboxing options, including dev containers, single-use virtual machines, and documented Mac OS sandboxing. To address security concerns, they implemented a mandatory package cooldown policy, preventing the installation of packages less than 7 days old across the company. This is enforced with Mobile Device Management (MDM).
Connecting Agents to Real Tools and Results [20:38]
With policies, guardrails, sandboxes, and skills in place, Trail of Bits connects AI agents to real tools, such as Slither, through an MCP server. This allows them to inject policy into how these tools are used. As a result, they have accumulated 94 separate plugins, 201 skills, 84 specialized agents, and 400 reference files encoding their domain expertise. This has led to significant improvements in bug detection, with some clients seeing an increase from 15 bugs per week to 200 bugs per week. AI now initially discovers 20% of all bugs reported to clients. On the business side, their sales team averages $8 million in revenue per representative, double the industry average.
Essential Starter Pack and Open Questions [23:20]
Guido summarizes the essential starter pack for becoming AI-native, emphasizing the importance of copying the system, not just the tools. He also discusses open questions and areas for future research, including private inference, prompt injection, and policy enforcement with continuous learning. They are exploring solutions like private inference servers, agent-native shells, and policy architectures to address these challenges. The business model is also shifting towards billing by expertise and results, rather than by hours.
References and Job Opportunities [26:22]
Guido provides links to the tools and references mentioned in the talk and announces that he open-sourced 10 additional plugins. He also mentions that Trail of Bits has job openings and encourages interested individuals to apply and help continue to advance their AI-native journey.
Q&A - Resistance and Blessed Lists [27:12]
In the Q&A session, Guido addresses concerns about job displacement, noting that Trail of Bits' senior-level workforce is well-positioned to benefit from AI enhancement. He also discusses the importance of maintaining a curated and verified list of approved plugins to prevent rug-pulling attacks and ensure the security of their AI tools. The curated skills repo is open source, with pull request reviews and automated systems for updating and verifying resources.
