Apiiro Turned a Security Disclosure Into Front-Page News by Refusing to Round a Single Number.

The Play

Apiiro's research into Claude Code's "Don't Ask Again" permission setting could have been written as a generic warning: "AI coding agents can leak secrets." That version gets ignored.

Instead, Apiiro published numbers that dared a reader to check them.

What the numbers actually said

The core claim was simple: when a Claude Code user selects "Yes, and don't ask again," the tool can save the full approved command — secret included — into a project file at .claude/settings.local.json.

What made it spread was the breakdown:

• "1,122 public copies" of that file "hold at least one secret"
• out of "104,185 copies" found across "103,617 repositories"
• including "379 database passwords," "43 SSH logins," "42" tokens for GitHub, GitLab, and npm, and "23 AWS keys"
• across "63 organizations," mostly software companies

Apiiro also compared 16 coding agents and reported that Claude Code was the only one it confirmed to save the full approved command by default — a direct, named comparison most disclosures skip.

Why this reads as research, not fear

Every figure in that list is a thing a reader can go check. The denominator (104,185 copies, 1.1% held secrets) does more work than the headline number alone — it tells the reader exactly how the risk is distributed instead of asking them to assume the worst.

That is the difference between a disclosure that gets cited by other outlets and one that gets one skeptical comment and disappears.

The growth play to steal

1. Name the exact file, setting, or artifact involved — never the category

2. Break a scary total into labeled subtypes instead of one round number

3. Publish the denominator alongside the headline count

4. Name every tool or competitor in the comparison, not "other tools"

5. Give a reader a way to verify the claim themselves in minutes

Why most teams skip this

Exact numbers are riskier to publish than vague ones — a reader can find the one case where your math is off, and a competitor can quote the number back at you forever. Rounding feels safer.

But that risk is also the whole mechanism. A claim specific enough to be checked is a claim specific enough to be trusted, and trust is what gets a disclosure picked up instead of scrolled past.

Bottom line

Apiiro's report didn't travel because coding agents leaking secrets was a surprising idea. It traveled because the numbers were exact enough to verify, broken down by credential type, and compared against 16 named competitors. That is a distribution strategy, not just a research result.

Source: https://thehackernews.com/2026/11/claude-codes-dont-ask-again-file.html