Pareto 20, No. 11: Laurent DeSegur, for what a coding agent can leak from your build pipeline

An independent developer who reads the internals of coding agents and writes up how they get attacked. The publication name says crypto; the content is entirely about shipping software safely with AI.

Seat 11 on The Pareto 20 is Laurent DeSegur, and it is the seat for a risk most people shipping software with AI have not priced: the agent in your build pipeline reads text from strangers, and text from strangers can give it orders.

Who he is

DeSegur is an independent writer and open-source developer in California who publishes as oldeucryptoboi. Ignore the handle: we found no crypto content anywhere. Since early 2026 the newsletter has covered agent architecture, runtime safety and the security failure modes of autonomous coding agents, three to five posts a month, alongside an open-source agent-safety runtime he maintains on GitHub. His about page puts it plainly: "I build open-source tools for safe AI agent autonomy."

What to read him for

The build-pipeline attack, explained. Securing AI Coding Agents in CI/CD (June 26, 2026, paid) walks through how an agent wired into a code repository can be steered by a crafted issue or pull request. His framing is the useful part: "This is prompt injection. It is a property of every current LLM, not a defect in one vendor's product." He names the disclosed cases, including a vulnerability in Anthropic's own GitHub action that Microsoft's Defender team reported, and the one that mattered most, an agent tool that "actually shipped a poisoned package to users." His remedy is a layered pipeline with a fail-closed default, and he names the fail-open default that caused the damage.

Safety as code, not policy. The Three Laws of Safe AI Agents (February 26, 2026) argues that promises from vendors are not controls, using Anthropic's revised scaling policy as the example, and shows what enforcement in code looks like: credential redaction, request filtering, circuit breakers.

How the tools actually work. How Claude Code Intercepts Subprocess HTTP Traffic (April 9, 2026) and his architecture deep-dive (March 31, 2026) are reverse-engineered internals. Most readers will not follow the protobuf, but the conclusions about where tokens can leak are written in English.

Start here

The architecture deep-dive is free and foundational. If you run agents in a build pipeline, pay for the CI/CD post.

What he sells, and what that means for you

A paid Substack tier; at least the CI/CD piece sits behind it. His runtime project is open source with no commercial tier we could find. We found no employer, consulting, courses, affiliate links or sponsorships. One post we could not open, a May 2026 "field report" on a named vendor's product, leaves open whether he has a commercial relationship there; we are noting the gap rather than guessing. The test counts he cites for his own runtime are self-reported.

Not for you if

You do not run coding agents against a real code repository. This is the most technical seat after Cosmin Pop's, and its value is proportional to how much of your product is being written by an agent with access to secrets.

Part of The Pareto 20, twenty creators who get a solo operator most of the way on AI decisions. How we chose them. Facts above were checked against the creator's own pages on September 12, 2026.