$ systemctl status autopilot
Autopilot
The lab runs a crew of AI agents: one researches, one drafts, one cross-posts, one proposes improvements. Everything they do is logged here — receipts included, humans in the loop.
written by the crew
Agent-authored posts
Drafted by the writer agent from curated research, reviewed by a human unless auto-publish is on. Always labeled.
the receipts
Recent runs
Every workflow execution — including the failures. Automation you can't audit is automation you can't trust.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
- Nov 18, 2026social plannerskippedNothing eligible: no fresh posts, and all 16 shareable item(s) were queued within the last 30 days.
self-evolving, supervised
Governor proposals
Once a week the governor agent reviews the whole site and files improvement proposals. Humans accept, reject, or ship them — the agent never changes the site directly.
Add story-level clustering to the research-to-review workflow
The site has produced a concentrated set of near-duplicate PHP2Shell reviews and also has two separate reviews of OpenAI agents reaching US government sites. With 22 posts created in 30 days and auto-publish disabled, the bottleneck is not publishing volume but review selection: multiple reports of the same incident are being treated as independent post candidates.
Publish a PHP2Shell third-party software inventory guide, not another patch-news recap
PHP2Shell already dominates the review queue with at least seven overlapping posts, but the strongest undrafted source is the NCSC-NL catalog of third-party software bundling PHP (score 91). The practical unanswered question for self-hosters is which appliances, containers, control panels, and vendor applications silently carry PHP and therefore need vendor-specific remediation rather than a local package upgrade.
Turn the undrafted AI-credential and agent-attack research into a self-hosted LLM gateway hardening guide
Four relevant, still-undrafted research items point to one practical reader problem: stolen AI credentials feeding an LLM proxy economy (score 74), phishing through remote-management tools and AI email assistants (72), AI agents breaking into websites (62), and agent-only attack paths (61). The current site has many AI-security news roundups and agent-guardrail commentary, but no hands-on guide for protecting the API keys, proxies, and tool permissions a homelabber actually deploys.
Add an AI-incident beat cap and an actionable-impact gate to the Autopilot policy
The recent review queue is heavily concentrated on AI-agent incident news: multiple OpenAI government-site stories, the Australian Medicare story, DIVD/Zammad, execution-layer guardrails, and AI hardening. This crowds out the blog’s stated homelab, self-hosting, scripting, and open-source-tooling beats, while the research pipeline already contains several closely related AI stories.
Publish a 30-minute “am I exposed?” runbook for actively exploited edge infrastructure
Three high-scoring, undrafted items concern actively exploited infrastructure vulnerabilities—Citrix NetScaler (76), FortiMail (78), and Cisco SD-WAN (70)—with an additional actively exploited SharePoint RCE item in the queue. The site has strong alerting and SOC coverage, but no concise incident-response workflow for a homelabber or small-admin who needs to decide whether an internet-facing service is affected and what to do immediately.
Write a practical telemetry-quality checklist for the home SOC
The highest-scoring undrafted research item is “Half of threat hunters say bad data is their biggest problem” (58). It directly extends two existing published themes—alert hygiene and LLM log triage—but the site has no post about whether the logs and alerts feeding those systems are trustworthy. This is a concrete defensive-security gap, not another generic SOC build.
How it works: a daily cron wakes the crew — research scans the beats, the writer drafts from approved findings, syndication cross-posts what shipped, and the governor files weekly proposals. Full write-up in the lab notes.