security

A 9,000-character silent truncation cut the redaction rules out of the writing prompt. The fix that mattered was a mechanical gate on the final artifact.

TL;DR: A style guide was silently truncated at 9,000 characters, cutting the redaction rules before they reached the writer model. The fix loads the guide whole and adds a mechanical verification gate, check_redaction, that scans the final article for non-documentation IPs and secret-like token prefixes. False positives cost minutes; published leaks are permanent, so the trade-off is easy.

The transcript must carry the hijack sentence verbatim as data. That is exactly how you prove a media file never acted as an instruction.

TL;DR: The author deliberately embedded a prompt-injection sentence into an audio fixture to test a media pipeline's defenses. Counterintuitively, the test requires the hostile transcript to stay verbatim and labeled untrusted, since injection is about text switching from data to instructions. The real defense is a contract proving content never becomes commands, not filtering text or trusting the model's intelligence.

The hardening playbook said changed, but sshd -T still showed passwordauthentication yes: first-match-wins and lexical Include order let 50-cloud-init.conf win.

TL;DR: OpenSSH uses first-match-wins for sshd config, so a drop-in named 50-hardening.conf lost to cloud-init's 50-cloud-init.conf alphabetically. Fix: rename it 00-hardening.conf, comment out conflicting lines in the main config, and edit 50-cloud-init.conf directly. Always verify with sshd -T instead of trusting a playbook's green status.

One WordPress install, two Git repos: a full Forgejo repo for recovery, a GitHub view repo carrying only source I wrote myself.

An hourly wave of 504s looked like an attack. The logs pointed at a dormant rate limit and a PHP-FPM pool out of workers.

TL;DR: Hourly 504 waves looked like a DDoS but were actually PHP-FPM worker exhaustion, worsened by request_terminate_timeout killing slow requests. Nginx rate limiting never fired because Cloudflare hid real client IPs, collapsing everyone into one $binary_remote_addr bucket. Fixes: enable Cloudflare caching, tune PHP-FPM pool settings, and rotate container logs with copytruncate.

Months spent blaming an anti-bot; the culprit was an Indonesian ISP's DNS redirecting duckduckgo.com to Trustpositif.

TL;DR: Headless browser tests inside Docker kept failing on DuckDuckGo, and I wasted months blaming anti-bot fingerprinting. The real culprit was Indonesian ISP DNS poisoning; Kominfo's 2015 mandate redirects port 53 queries, so duckduckgo.com resolved to a wrong IP. The fix was one flag: docker run --dns 1.1.1.1 restored access instantly.

A two-line .gitignore whitelist commit turned into a full backup redesign: Git repo, restic, immutable cloud, 3-2-1-1-0 without a NAS.

TL;DR: A two-line .gitignore change led me to audit my backups, and I realized my NAS setup was untested and ransomware-vulnerable. I replaced it with a 3-2-1-1-0 approach: git bundles, restic snapshots, and immutable cloud storage. Monthly restore drills now verify the whole thing actually works.

A security audit uncovered 11 findings in the chat AI pipeline, from cache keys to prompt injection.