The Security Handcraft: 12 Fossil Scripts, One Live Key, Zero Leaks in Print
Key Takeaways — Executive & AI Summary
- The rule as built: one credentials home per machine — the workstation config holds 138 fields of which 6 are credentials, file mode 600 — read at runtime by code like the image factory, with a logging filter that swaps token substrings for asterisks before any line is written. New code never inlines.
- The rule as inherited: the audit found 12 legacy push scripts each carrying the same 46-character token inline, and a live probe returned ok=true for both that fossil string and the config's string — 2 valid credentials for one bot, only one of which the config knows about.
- The publication gate holds: 0 secret-pattern hits across all 23 published posts, 5 ritual-scan receipts in the work log, and the two in-session forensic leaks this series caused (a config dump, a parent-object print) were both caught before print.
Episode 23 of One Man One Legion — the fifth war story. Episode 21 autopsied a framework that died in ninety minutes; this one turns the fleet’s own leak scanner around and points it at the machines.
The ritual
Every episode of this series ends the same way. Before anything is pushed, both files — article and ledger — are pattern-scanned for the fleet’s red-line items: token shapes, group ids, ports, proxy addresses. The work log carries 5 receipts of that single line being written. And today’s whole-library run covered all 23 published posts in the repository: 0 hits for Telegram-token, sk-, ghp-, or AKIA-shaped strings.
This episode asks the obvious next question. What happens if you turn that scanner around — away from the articles, onto the fleet itself?
Where the keys live
The rule as built: one credentials home per machine, read at runtime, never inlined. On the workstation, that home is a single config file of 138 fields, of which 6 are credential-shaped — 3 apiKey entries, 1 appSecret, 1 botToken, 1 token — sitting behind mode 600, owner-only. Modern code doesn’t copy from it; the image factory’s helper function loads the bot token fresh on every push and skips quietly if the read fails.
The AWS orchestrator repeats the pattern with more machinery: a dedicated secrets file, mode 600 verified today, environment-first with file fallback, and a logging filter that replaces any token substring with asterisks before a line reaches a file or a console — its header calls this constitution clause 4 — keys never enter logs. The account’s two inference environment variables live in a shell profile by name only.
How keys move
Never in a command line, never across the wire. One incident set the doctrine. On Aug 11, during the fleet-monitor build told in Episode 5, a bot token appeared inside a conversation. The work log’s warning line recorded the exposure and prescribed the response in the same breath: the human re-issues the token at the bot authority, then places the new string into the 600-permission file by hand over ssh — a replacement credential never crosses a chat again. That doctrine is why this episode’s audit compared hashes, never strings.
The audit that found something
A pattern scan across the workstation’s code matched secret-shaped strings in 16 files. 4 of those were upstream test fixtures — a packaged retrieval project ships fake keys in its own test suite; a scan’s hits are not all yours. The other 12 were ours: the legacy push lane. News-push v3, v4, and v5, two hotspot-monitor skills, two social-trend monitors, a reddit monitor, and video-batch helpers — every one of them carrying an identical 46-character token string, byte for byte. The single-home rule arrived after these scripts were written, and nobody went back.
The fossil was supposed to be dead — supposed to be. No record this series has ever read tested that. Today’s getMe probe, run on the workstation through its own proxy so neither string left the machine: the config’s token answered ok=true, and so did the fossil — same bot, 2 distinct live strings. The config only knows about one of them. A revoke kills a string, and a live string is its own proof: no revoke ever reached this lane, though one was recommended once, on Aug 11, for a different leaked token. The fossil authenticated fine today.
Why it survived is also why the cleanup is finally cheap: since Episode 19’s contraction, no scheduled job invokes any of the 12. They are fossils with a pulse — dark cron-wise, alive credential-wise. The remediation order writes itself: inventory, migrate or delete, revoke, rescan.
The leaks that didn’t reach print
Twice, this series’ own evidence-gathering printed a secret into a tool session — a config dump that carried an appSecret during the pen-name audit, a parent-object serialization that carried an apiKey during the brain-swap dig (Episode 10, Episode 18). Both were caught in session, neither entered print, and both final files scanned clean. The recurring lesson is uncomfortable: the most likely leak tool in a transparent infrastructure is your own forensic code.
Posture, briefly
The network side is told elsewhere and hasn’t changed: public SSH closed since Aug 9, render submissions gated behind a two-entry port whitelist with path checks. The one CVE war story — a torch.load security gate whose instructed upgrade produced five verbatim breakages in one morning — was published separately and stands where it stood.
Security in this fleet is not a product that was bought. It is a habit of scanning itself, the same grep that keeps 23 published posts clean having just found a live fossil on the machines. Transparency is paid for in redaction — and occasionally redeemed by it.
Back to the fleet audit — one man, one legion, and the discipline that lets it publish its own logs.
FAQ — Direct Answers
- Doesn't publishing this point attackers straight at a live key?
- No string, bot name, group id, port, or proxy address appears — the same red lines every episode is scanned against. The machine sits behind Tailscale with public SSH closed since Aug 9. And the finding's value is the method: pattern-scan, hash-compare, verify with an on-box getMe so the credential never leaves its host.
- How should the fossil lane actually be cleaned up?
- Inventory, migrate or delete, revoke, rescan. No scheduled job has invoked any of the 12 scripts since the Episode 19 contraction, so revocation breaks nothing that is lit — the only residual risk is undocumented manual use, which the inventory step exists to catch.
- Why compare token strings by hash instead of just diffing them?
- So the comparison itself never copies a secret into a transcript or log. A digest answers same-or-different without revealing the value — the same principle as the logging filter that masks tokens before a line is written anywhere.
- What do the scans actually cover?
- Known shapes: Telegram bot-token format, sk-, ghp-, and AKIA-style strings, plus this fleet's own red-line items — group ids, ports, proxy addresses, voice names. Custom-format keys are not covered. The audit is a habit, not a guarantee; that is why it reruns every episode instead of claiming to be done.