Loading live data…
Security operations,
engineered end to end.
A Kubernetes GPU cluster, defended in layers and watched around the clock. Every number on this page comes from the environment's own sensors, sanitized and published one way, an hour late, on purpose.
Data as of
—
—
IDS sensors
2 / 2
North–south TAP · east–west SPAN
Hosts reporting
—
Host agents checked in, last 24h
Guardrails tested
1 / 10
Building toward 10 / 10
Last guardrail test
Nov 20
2026 · all checks passed
External threat activity
Unsolicited connection attempts against the environment's public address · last 24 hours—
probes blocked
—
unique sources
—
serious host alerts
Blocked attempts per hour
What they were looking for
Where the scans came from
Location shows where scanning infrastructure is hosted, not who operates it.
Defense in depth
Each layer catches what the one before it lets through01
Perimeter firewall
Sophos XG · zones, IPS, default deny
—
02
Network intrusion detection
Suricata · TAP + SPAN, ~53k rules
—
03
Host monitoring
Wazuh · file integrity, audit logs
—
04
Workload guardrails
Kubernetes · Pod Security, default-deny network
Policy enforced
AI guardrails, with evidence
No model runs until every guardrail is built and testedTested prevented, detected, proven Built control in place Planned
AI-01Built
No data leaves
Default-deny egress; cloud AI APIs blocked and alerted
AI-02Planned
Prompt-injection awareness
Hostile input flagged before an agent reads it
AI-03Planned
Agents stay read-only
Read-only credentials; any write is alerted
AI-04Planned
Supply-chain integrity
Pinned images; model files watched for change
AI-05Tested
No breakout or tampering
Pod Security enforced; exec, secrets and RBAC changes audited
AI-06Planned
Model API is private
Only the agent may reach the model
AI-07Planned
GPUs used only by the model
Unexpected GPU processes alerted
AI-08Planned
Every agent action logged
Tool calls and results recorded centrally
AI-09Planned
Kill switch works
One command stops every agent; drilled monthly
AI-10Planned
One-way, sanitized publishing
This page: aggregates only, no inbound path
Engineering log
Real decisions, and why they were made- Published this page through a one-way pipelineHourly summaries leave through a write-only drop box; the reporting account can read nothing but those summaries, proven by test.
- Locked down the AI workspace before any model runsPod Security and default-deny networking in place; breakout attempts are rejected and audited.
- Built an alert for log feeds that go quietA feed failed silently after a nightly rotation. Now silence itself raises an alarm.
- Separated signal from noise in firewall telemetryBroadcast chatter is dropped at ingest; probes against the firewall itself are kept and analyzed.
- Fixed container traffic split across sensor threadsDetections between containers went from 0 of 3 to 3 of 3, proven by test.
How this site is secured
This page never talks to the environment it reports on. Sanitized summaries are pushed out; nothing on the internet can connect in.- No inbound connections to the environment
- Aggregates only: no IPs, hostnames or versions
- No vulnerability details published
- Hourly roll-ups, delayed by design
- Static site: no database, no server-side code, no third-party scripts
- Write-only publishing credentials, read-only reporting account