Dan Guido, CEO and cofounder of Trail of Bits, has described how his firm moved from AI‑assisted tools to an AI‑native operating model. Rather than treating AI as a one‑off procurement, Trail of Bits built a repeatable system: a visible capability ladder, recurring hackathons, curated skills repositories, hardened defaults and sandboxing to address human resistance, data permissions debt and supply‑chain risk.
Why this matters now
A Fortune report in February summarized a National Bureau of Economic Research study in which nearly 90% of roughly 6,000 executives said AI had produced no measurable change in employment or productivity at their firms over three years. Guido argues this isn’t evidence AI doesn’t work—it's evidence most companies deploy AI incorrectly by handing out licenses and waiting for magic.
Three levels of AI adoption
Guido distinguishes three adoption levels:
- AI assisted: basic productivity helpers (e.g., ChatGPT drafting emails or summarizing documents) without changing workflows.
- AI augmented: workflows are redesigned so AI does first passes and humans do oversight.
- AI native: structural redesign that treats AI as a core participant or teammate rather than a tool.
Trail of Bits aims for expertise to "compound as code"—every engagement should produce reusable artifacts that make the next engagement faster and better.
The first barrier: employee resistance
When Guido announced last year that the company would integrate AI across workflows, roughly 5% of employees immediately supported it and 95% resisted; about 20% resisted actively and the rest passively. Instead of confronting people, Guido studied why people reject new technology and identified four biases to address: self‑enhancing bias, identity threat, opacity and intolerance for imperfection.
- Self‑enhancing bias: people credit successes to their judgment and blame losses on circumstance, common among senior staff.
- Identity threat: knowledge work can be symbolic of who someone is; framing AI as doing the job triggers rejection, but framing AI as making the person a better practitioner is accepted.
- Opacity: lack of visibility into how decisions are made by models.
- Intolerance for imperfection: people abandon algorithms after seeing them err once, even if they outperform humans overall.
Guido designed specific countermeasures for each bias.
What Trail of Bits built
-
Capability matrix (AI maturity matrix): a visible four‑level ladder specific to AI skills, separated by role (assurance, engineering, sales, project management). This prevents the "we’re already good enough" attitude.
-
Skills repositories: three collections—a private internal repo for company workflows, a public repo for external reuse, and a curated marketplace that vets third‑party skills to manage supply‑chain risk.
-
Hackathons: recurring, goal‑oriented events every two months (roughly) with a defined focus and learning objectives, pairing participants so all work is reviewed, demos and follow‑through. The metric is participant movement on the capability ladder, not only shipped artifacts.
-
"Scar tissue" turned into defaults: repeated failures or gotchas are encoded into global, copy‑pasteable defaults (e.g., claude‑code‑config) so new hires don’t need to relearn lessons.
-
Sandboxing and hardened defaults: multiple isolation options (devcontainer for devs, dropkit for disposable droplets, COOP for isolated VMs, plus Claude Code sandbox). They also enforce procedural defaults such as a seven‑day delay on every package their developers install to let security research pick up malicious packages first.
-
AI handbook: a written usage policy and risk model so that risk decisions aren’t ad hoc.
Operational details and numbers
- Trail of Bits is about a 130‑person company.
- They tracked roughly 50 engineering skills for performance reviews and moved AI skills into a separate four‑level matrix: not engaged (level 0), capable, adoptive and transformative (level 3).
- The highest maturity level is not the most frequent AI user but the person who invents new ways to work and builds AI‑based tools.
- Hackathons run roughly every two months; after each one 1–2 people harvest and structure reusable artifacts.
- They maintain three skills repositories: internal, public and curated.
- A seven‑day cooldown is applied to all package installs.
Data and DJ Patil’s "Tidy House"
Guido says the most common blocker when building AI workflows is data and permissions: who can access which systems and where data lives. Agents make permissions debt visible; organizations must audit permissions, decide boundaries, and make data reusable to compound value. Guido prefers a single centralized foundation that teams build on top of rather than many uncoordinated solutions. He recommends measuring what fraction of a team’s data is reusable and consumed by other teams.
What post‑AI jobs look like
Trail of Bits taught non‑technical staff basics of git and the command line so they understand what happens under the hood. Many routine tasks will be automated, but new roles appear: quality assurance for agent systems and skills product management, a role the firm already staffs. Tools from hackathons enabled technical editors to stop line editing and instead consult on framing and voice.
Measurements, telemetry and benchmarks
Telemetry from developers’ dotfiles and device management shows what skills are used and what breaks. An AI systems engineer serves as product manager for the skills repository—reviewing pull requests and deprecating overlapping skills. Trail of Bits is now building benchmarks for core skills (e.g., how well an agent finds bugs in a language or produces a statement of work) to give evaluative proof during performance reviews.
One mistake and the gym analogy
Guido’s main regret: not allocating enough FAFO time ("F Around and Find Out") early on—trying new products, documenting them and systematizing knowledge quickly rather than waiting for others to do it. He compares skill acquisition to going to the gym: you need reps.
The recipe
Guido summarizes his repeatable playbook as:
- Standardize on one agent workflow you can support.
- Write an AI handbook so risk decisions are consistent.
- Create a capability ladder that makes improvement expected.
- Run short adoption sprints (hackathons) for hands‑on practice.
- Capture everything as reusable artifacts: skills + configs + a curated supply chain.
- Make autonomous agents safe with sandboxing, guardrails and hardened defaults.
Trail of Bits has published its skills repository, curated marketplace, configuration repo, devcontainer, dropkit and COOP; the playbook and slide deck are available on their blog and GitHub.
Conclusion
Trail of Bits’ approach shows that realizing AI’s productivity gains is primarily an organizational problem: incentives, defaults and measurable mechanisms must be designed so AI adoption diffuses across the company. The Solow paradox resolved in the past because firms reorganized around computing; Guido argues the same kind of organizational work must now happen for AI, not just technical advancement.
