Somebody still has to read it

A GitHub repo created a few days ago published more than fifty CVEs. NVD flagged the SQLite ones critical. CISA's ADP agreed. Red Hat scored one of them a 10.0.

Then JFrog looked. The cited functions weren't in the versions named — `exprComputeOperands()` was added to SQLite in the middle of 2025, and the advisory pinned the bug to 3.41.0. The proof-of-concept payloads ran clean under AddressSanitizer. None of the CVEs appeared on SQLite's own advisory page. Fed through GPTZero, the advisories trip the AI-generated warning. Red Hat's 10.0 is now a 7.6.

Here is what establishing that cost: clone the repository, check out three version tags, compile each release in an isolated Docker container, run every PoC under ASan, then audit the CPE metadata across the NVD and GHSA feeds. Here is what producing the claims cost: a prompt. That ratio is the story. Not the slop — the slop is boring, and there will be more of it tomorrow. The ratio.

The same asymmetry surfaced twice more on Hacker News today, in places with nothing to do with security. The day's top post is a complaint about Slack: ask a question, get back a model's answer pasted verbatim. Reading it is work. It is verbose, jargon-dense, and, in the author's phrase, frequently contains all too plausible nonsense. His ask is small.

Then he follows it to where it actually bites. Paste the ticket into Claude Code. Don't open the diff. Paste the reviewers' feedback back in. Iterate until it merges. It works — and "who has done the implementation? The reviewers did, using Claude Code, and you as a meat proxy." The commit carries your name. The reading was done by someone else.

Ankur Sethi's answer is to pay the cost himself, on purpose. He has the assistant write code in the chat window, then types every line into his editor by hand. He calls the workflow "grossly inefficient and perhaps slightly comical." He figures it drops him from 10x to 2x. He does it anyway, because the typing is what builds the map of where things live — and because a codebase nobody understands is, in his words, professional malpractice.

The obvious objection is that this is spam, and spam is a solved problem: rate-limit it, reputation-score the submitter, filter it at the door. That works when the filter is hunting for garbage. None of this is garbage. The advisories are well-formed — real function names, plausible use-after-free mechanics, correctly structured CVSS vectors. The Slack paste is articulate. The unreviewed PR passes CI. What broke isn't quality control. It's that every one of these systems assumed a credible-looking claim would be expensive to make.

What broke isn't quality control. It's that every one of these systems assumed a credible-looking claim would be expensive to make.

NVD's job was never independent verification; it enriches and scores what gets published to it. That was adequate for as long as fabricating an advisory cost more than dismissing one. It does not survive a week-old repo shipping fifty at once, and the same inversion is running in the pull request queue and the Slack thread. The work didn't disappear. It moved downstream, onto whoever is still in the habit of looking.

One developer trades 10x for 2x and calls the difference comprehension. The other asks only for a sentence in your own words. Both are paying by hand for something that used to come bundled with doing the work. Somebody still has to read it. The open question is who — and whether they know they've been volunteered.