JBNX daily CEO report — 2026-08-12

JBNX daily — Wednesday 12 August 2026

BLUF: The security remediation pass landed — ten production fixes applied and verified across FedM8, CipherDeck Apps, billing and the directive, and the blocked-work queue went 6 in, 6 out and is empty tonight. The thing needing the CEO: the weekly attack-surface audit found a P1 — test.projects.jbnx.io serves the ungated write API against the production database with its own rate-limit bucket, so an outsider can forge claims and billable sessions with no credential.

FOUND

Security

• P1 — test.projects.jbnx.io is the same Railway service as production, publicly reachable, unauthenticated, reading live production data, and its rate limiter is counted per host. Alternating the two hostnames yields 2x the intended write rate on /api/1099/. Downstream: forged claims that lock the fleet out of every slug, forged billing.work_sessions a customer is invoiced for, forged handovers and /framework/ documents. (/framework/attack-surface-2026-08-12)
• P2 — mkt.jbnx.io: TLS hostname mismatch, and nothing behind it.
• P3 — GitHub Pages hosts serve no security headers; fedm8.com returns wildcard CORS.
• P4 — bill.jbnx.io is missing Permissions-Policy.
• GITHUB_TOKEN is present in a production Cloudflare worker. (/framework/security-review-2026-08-12 v2)
• The billing schema is unlocked — found during the Dispatch v2 gap review, outside the original audit scope.

Infra

• Five scheduled agents due this morning did not run: docs-directive-audit (5:49), agent-fleet-audit (6:03), fedm8-sales-feature-brief (6:09), uptime-alert-sweep (6:24), code-audit (7:30, Wednesday rotation). No billable session, no document, and no scheduler last-run timestamp for any of them. First estate activity of the day was 07:47. This is a silent fleet failure, not a thin day.

Cost / governance

• Measured against 14 days of live activity: 19% of AI work ran outside any project claim (unbilled), and 49% of claims produced no artifact. (/framework/dispatch-control-plane v2)
• 50 claim collisions today, clustering on projects-portal and the jbnx.io, claude-ops and fedm8-scan repos — repeat collisions on one slug mean the work is badly split, not that agents are misbehaving.

FIXED

• ops_is_admin() changed from fail-open to fail-closed on both the FedM8 and CipherDeck Apps databases. Verified 3/3 probes each.
• 16 dead write grants revoked on FedM8; two trigger functions and ops_is_admin removed from the anonymous RPC surface (trigger still fires).
• ops_log_404 user-id spoofing closed; client IP now derived from headers. Spoof attempt rejected on probe.
• admin_* RPCs removed from the anonymous surface on CipherDeck Apps.
• 78 orphaned billing sessions backfilled and 12 unassigned projects assigned to the paying customer; a trigger added so orphans cannot recur. 0 remaining, probe passed.
• Three live exploitable Supabase advisor findings fixed: portal view disclosure, paywall stub, anonymous request amplification.
• hi.jbnx.io single point of failure repaired — server.js wanted sql/directive-v35.md, which did not exist on main. Now at v42, byte-identical, all four endpoints agree.
• robots.txt on projects.jbnx.io now names ClaudeBot and Claude-User while keeping Disallow: / for the wildcard. Commit 215b1f24. Zero failed uptime checks since; it had been producing 7 false DOWN alarms per 24h.
• CipherDeck: PR #47 merged as ce41af9c at 18:55Z after GitHub Actions billing cleared; PR #48 (button outage — production CSP now permits the Supabase auth origins) merged as ea5294c8a83d, Railway production SUCCESS, live at app.cipherdeck.com.
• agent.sh gained the usage subcommand the operating rules already required; shipped to main.
• FedM8: 11 files shipped to fedm8-scan main, all six hi.fedm8.com surfaces return 200 and agree at v8; performance fixes (sized WebP logos) live on ai.fedm8.com.
• Dispatch Phase 1 applied to production: a registry of approved AI worker identities with per-actor tokens.

The hourly resumer cleared all six blocked items on its own today, including the CipherDeck GitHub-billing deploy block, without anyone touching it.

FINISHED

Work sessions closed today, in their own words:

• Confirmed the CipherDeck product improvements are live for customers; the earlier hold-up, a GitHub billing problem that stopped the automated quality check, is cleared.
• Button outage root cause fixed: production now permits the sign-in service it needs, and the Add controls are guarded.
• Repaired the single point of failure on the master agent directive page and added the token-usage command the operating rules already required.
• Stopped the false site-down alerts caused by our monitoring being blocked from reading the site, without exposing the internal portal to search engines.
• Checked every website and service the company exposes to the internet the way an outside stranger would see them — 25 addresses.
• The FedM8 site now tells search engines what it sells and what it costs, and the money-back guarantee is visible where people decide to buy.
• Measured 14 days of agent activity against the live system and published what it showed.

81 work sessions across 13 slugs, 80 handover updates across 9 projects, 23 documents published. Four sessions still open at 4pm — none older than 12 hours: cipherdeck-apps (2), projects-portal (1, Dispatch Phase 2/3 build-out), and this report.

BLOCKED

Nothing. Six items entered the queue today and all six resolved the same day. Longest wait 1h10m (the hi.jbnx.io fallback, waiting on a projects-portal claim); shortest 5 minutes. Two were claim contention, one the GitHub Actions billing block, one a chat approval. Nothing needs abandoning, nothing is over 72 hours, and the queue is empty tonight. On this measure throughput is not degrading.

RECOMMENDED ACTIONS

AWAITING YOU

1. Decide the P1 test-lane fix. Reject non-GET /api/1099/* when the Host is test.projects.jbnx.io (the test lane only needs the SPA and reads), or key the rate limiter on client IP alone. Cost: roughly one agent hour. If it waits, the only compensating control on a deliberately ungated write API stays halved, and forged billables are invoices you would have to unwind.
2. Rotate the GITHUB_TOKEN sitting in a production worker. Only you hold that credential; no amount of agent retrying clears it.
3. Approve pushing the three written-but-undeployed remediation items: version-agnostic directive plumbing in server.js, plus secret and dependency scanning across 15 repos. They exist locally with 16/16 coverage; they need a PR and your chat approve.
4. Decide what to do about the five no-show scheduled agents. Either the scheduler needs repair or those agents should be cut — right now you are trusting a morning sweep that did not happen.

THEN

1. Set a daily model-spend ceiling. Today cost $1,323 in recorded model spend to produce 12.6 billable hours. That ratio is defensible on a remediation day and not on a normal one; without a cap you will not find out which kind of day it was until the invoice.
2. Split the projects-portal slug, or scope claims by directory. 50 collisions in one day is a work-breakdown problem — agents queued behind each other for the same lease three separate times today.
3. Decide on Dispatch Phase 2/3, currently mid-build under an open lease. It is the structural answer to items 5 and 6 above; leaving it half-applied is worse than either finishing it or stopping it.

SPEND

• Model spend today: $1,323.12 recorded, 69.1M input / 4.0M output tokens across 155 usage events. Caveat: 111 of those events recorded $0 (LiteLLM, Cursor and ChatGPT paths are not costed at source), so the true figure is higher than the recorded one.
• Billable time recorded: 45,437 seconds = 75,728 credits = $757.28, across 81 sessions and 13 slugs.
• Automation fleet's own cost: $11.92 across 8 usage events by scheduled actors — under 1% of the day's model spend. The fleet is cheap; the interactive chats are the spend.

---

Email sent to x@jbnx.io, message id 3055b8a5-a29b-470b-8ccd-f2a1ef3ad4c5. Token usage: in=192,000 out=15,000 model=claude-opus-5 engine=claude-cowork · recorded (estimated) Token health: 13:1 · Neutral → normal for chat; finish the task