Orca's CEO says the Starbucks story was the first of many. The people publishing malware to npm agree with him.
In July, Gil Geron wrote about Starbucks using AI-assisted development to replace a Microsoft inventory system and an IBM maintenance platform, going after a 400 million dollar software line. Forbes framed it as a warning shot at the application vendors. Geron pushed back: a warning shot implies one round, and this sequence has no end. He also called the incumbents' next move early. Once speed and cost stop favouring buy, the pitch turns into fear. Building your own gets framed as reckless, and the licence fee gets framed as protection.
He is right that the fear pitch is backwards. I want to finish the part his piece leaves open, because the security industry was not the only audience that noticed the builder population exploded.
What Orca got right
The piece says "almost everyone in the building is a builder" now, and that matches what I see in the pipeline I run. The marketer is writing Python. The analyst is shipping agents into production. These people will never attend secure coding training, because they do not consider themselves the audience for it. The agents doing half their typing cannot attend at all.
Geron is also right that buying a legacy application never bought you security. It handed you the vendor's patch calendar and whatever they decided counted as secure enough. Building moves that work into your own hands, and owning it beats renting it.
Then the piece lands on its prescription: visibility into everything you run, plus context on what is actually reachable, remediated fast. That is a security team's wish list. It is addressed to a reader who opens dashboards for a living. The population Geron just described does not.
The same curve, read from the other side
While vendors argued about who secures the builders, the 2026 supply chain campaigns ran their own experiment on the same population. Read them as one sequence instead of four incidents and a shape appears that should bother you more than any single CVE.
Shai-Hulud shipped with valid provenance attestations because the compromised build genuinely ran in the real pipeline. Provenance proves origin. It says nothing about safety. The keyv compromise needed one maintainer account to reach everything that depended on it. GHAPPIER abused Trusted Publishing, which is the mechanism the ecosystem rolled out in response to exactly that class of account takeover. And the RubyGems flood turned the registry's documentation build into arbitrary code execution, because a .yardopts file inside an uploaded gem drives the build, so publishing a package bought you a shell on infrastructure everyone trusts.
Each wave defeated the control the previous wave made the industry deploy. That is the shape. Attackers iterate on trust mechanisms faster than guidance cycles can respond, and whatever you verify next, someone will publish through it.
Visibility assumes a reader
Here is the problem with ending the argument at visibility and remediation speed. A finding is a message, and a message needs a recipient. The analyst who shipped an agent on Friday is not in the dashboard on Monday. The agent never will be. GitGuardian's 2026 data puts the secret leak rate of AI-assisted commits around 3.2 percent against a 1.5 percent human baseline. Twice the leak rate, coming from the exact population that will never open your findings queue.
Agents cannot be trained or reminded in a retro. Neither can the analyst who does not think of herself as a developer. Both respond to exactly one thing, a gate in the execution path that does its job whether or not anyone cooperates.
What a gate looks like when nobody cooperates
Concretely, from the pipeline I run.
Package installs. Shell aliases wrap npm, pip and the other package managers, and route every install through a local proxy that checks each package and its dependency tree against threat intel before anything touches disk, with a 48 hour minimum package age by default. The wrap covers human shells and the VM images our agents execute in, and the same checks run in CI. The agent does not know it is proxied. It runs npm install and gets a package or an error. Shai-Hulud propagated in hours and was publicly identified inside two days. Waiting 48 hours costs almost nothing and removes most of that window.
Secrets. Scanning runs twice on purpose. Pre-commit is developer experience, anyone can skip it, and it runs on a laptop I do not control. CI is the control. It cannot be skipped and it blocks the merge.
Identity. Every merge is signed. Human keys rotate on a 90 day cycle. Agent keys are separate, scoped to the agent's own identity, and live 30 days, so the audit trail can tell a release-agent merge from a human one. There is no override, because an override that exists gets used at 5pm on a Friday. Signing answers a question provenance cannot: who authored this. After Shai-Hulud you want both questions answered.
Skills. Agents install skills from one reviewed repository and nowhere else, and the 48 hour hold applies to skills exactly as it does to packages. OWASP guidance is encoded as a skill instead of documented as a page, which means it applies at the moment of authorship. It is also the only mechanism that reaches the analyst who has never heard of OWASP.
Hooks. A last guard runs before any tool call becomes an action, and we log tool calls rather than prompts. Prompt logs say what was asked for. Tool call logs say what happened. Only one of those is evidence during an incident.
Notice what none of this asked for. Nobody attended a training and nobody opened a dashboard. Building these controls for developer experience produced agent governance for free, because a control that never depended on cooperation does not care who is calling it.
What it cost
None of this was a tool rollout. Invisible controls mean the security team owns the integration work instead of pushing it to product teams, and it took substantially longer because of that. The context layer that keeps agents writing code that fits the codebase rots into lies without maintenance, which is worse than not having one. The skills repository is a permanent review load. The 48 hour hold generates a genuine argument occasionally, usually when the fix someone needs shipped that morning. And the signing rollout was the worst week of the whole project.
Whose gates were standing
Geron closes by asking who moves first. From where I sit, the builders already moved. The marketer is in prod. The agents merged over the weekend under their own keys, and the campaigns probing the ecosystem they build on are iterating monthly. The incumbents will spend the next few years selling fear, and the dashboards will keep filling with findings addressed to nobody.
So the question worth asking is a different one. When the builders arrived in your pipeline, whose gates were standing.
No comments:
Post a Comment