Detection · Defence · Offence · One loop
Stop bracing for the AI hacker.
Hit them at the source first.
Attackers now run machine-paced recon against your home computer: hundreds of probes in seconds, chosen by a model, not a person. Your antivirus waits to be hit. RIFM is the first AI driven offensive weapon that closes every surface, profiles whoever knocked, and pushes that attacker back into the swarm before the next wave lands. You buy a licence, you run one command in a terminal, your house becomes a node.
01 · Why this, why now
The attacker stopped being a person. Your defence did not notice.
An AI hacker does not sleep, does not get bored and does not guess. It enumerates your house the way RIFM enumerates a target: the full tree, in parallel, in seconds, and it learns which ports answer. A human analyst gives up at branch four. A model does not. That asymmetry is the whole problem, and it is why detection alone is a losing position.
Machine-paced recon
41 probes in six seconds is not a person poking at you. It is a model deciding what to ask next, and it will keep asking while you sleep.
Your footprint is the map
Handle, avatar, geotag, reused password, that old dev domain. Everything RIFM can find on a stranger, an attacker can find on you, for free.
Reactive tools
Signature scanners answer after the fact. They tell you what already happened, in a log you will never read, on a machine that is already gone.
No memory between houses
The same wave hits a thousand homes in sequence and each one learns the lesson alone. Nothing is shared, so everyone pays the same tuition.
02 · The product
One intelligence loop, four moves, no exit
This is the thing you can sell: not a scanner, a loop. Every home that joins makes the next home harder to breach, and every attack that fails becomes a signature that never works again. The value compounds, which is why the licence is recurring and why the first hundred nodes are worth more than the thousandth.
03 · Home shield
Everything aspect of the house, armed by default
Not a checklist you have to work through. The agent binds all fourteen surfaces on install and reports the posture as one number. Twelve of them are shown here, and two are in a degraded state, because that is the honest picture of a real home machine on day one.
04 · Delivery
They buy it, they paste one line, they are armed
No console to learn, no dashboard to log into, no agent to babysit. The licence email carries a single command. It runs in the terminal on their own machine, binds to that host, and reports posture in one screen. Click the console below, or the command chips, and it runs exactly as it would on their computer.
What the licence email contains
| Line 1 | curl -fsSL rifm.sh | sudo bash |
| Line 2 | rifm --activate RIFM-XXXXXXX-HOMELAB |
| Host binding | key is bound to the machine |
| Time to armed | under 90 seconds |
| Data uploaded | hashed signatures only |
| Uninstall | rifm --disarm, removes its own rules |
Sell it as a product, not a subscription you have to justify every month: the buyer sees a blocked-probe counter, a posture score and a strike log within a minute of install.
05 · Live lab
The offensive engine, running in your browser
This is the same detection half of the loop that runs against your own footprint. Pick a target type, run it, and read the findings back out loud. Seeded dataset, client side, nothing leaves the page.
Passive sources only. Nothing here leaves your browser.
06 · Threat model
What lands on a home computer, and what it costs
Severity is what it costs the household, not how hard it was to find. Detection confidence and impact are separate columns on purpose, because a high-confidence finding can still be a low-stakes one.
| Finding | Loop move | Severity | Detection conf. | Median time to exploit |
|---|---|---|---|---|
| AI recon sweep, 41 probes in 6s | Detect | Critical | 0.96 | minutes |
| Reused cleartext password, 41 months | Defend | Critical | 0.97 | hours |
| Open remote service with a live banner | Defend | High | 0.95 | minutes |
| Geotagged post at the home address | Attack | High | 0.74 | immediate |
| Unpatched package with a live exploit | Defend | High | 0.88 | days |
| Attacker tooling fingerprinted | Attack | Advantage | 0.91 | blocks next wave |
| Telemetry above policy, 3 vendors | Learn | Medium | 0.82 | n/a |
| Registrar details publicly readable | Defend | Low | 0.99 | n/a |
07 · The engine underneath
Async agent orchestrator, 140 concurrent tasks on one event loop
A `while True` agent loop blocks the main thread, so one slow lookup stalls every other agent you are running. RIFM does not work that way. FastAPI plus an asyncio event loop, every agent a coroutine, every tool call awaited and time boxed, bounded retries with jittered backoff, and a circuit breaker that drops a dead source and reroutes to a passive mirror. The run still finishes, cited, when half the internet is on fire. Everything below is the real server, live.
Fire a run
POST /runs · GET /runs/<id> · GET /stats · GET /tools · GET /events (SSE)
Tool health, straight off GET /tools
| Tool | Calls | OK | Failed | Retries | Fallbacks | p50 | Breaker |
|---|
Event driven, not blocking
Every agent is a coroutine. A 2.4 s LLM hop costs nothing but a suspension, so 140 branches share one thread and one loop instead of 140 stuck threads.
Timeouts on every call
Each tool has its own budget, from 1.2 s for DNS to 4 s for the model. A hung upstream is cancelled, not waited on, and the failure is recorded as a timeout.
Retry with jittered backoff
Bounded attempts, exponential delay times a random factor, so a fleet of agents recovering together does not stampede the source that just came back.
Circuit breaker, then degrade
Three consecutive failures open the breaker and the branch reroutes to a passive mirror. The run returns partial and cited rather than throwing.
Cancellation is clean
Disarm the house and the task tree is cancelled, semaphores released, no orphaned sockets. Graceful shutdown is a first class path, not an accident.
Streamed, never polled
Agent events go out over server sent events from an asyncio queue that drops its oldest frame instead of blocking the loop on a slow reader.
Why while True loses
| Blocking loop | Async orchestrator | |
|---|---|---|
| 140 agents | 140 threads, or one queue | 140 coroutines, one loop |
| Slow tool | stalls the whole loop | suspends one task |
| Dead source | raises, run dies | breaker, fallback, partial |
| Shutdown | kill the process | cancel tree, release locks |
| Observed | n/a | 1,067 tool calls in 13.2 s |
Run it yourself
pip install fastapi "uvicorn[standard]"
uvicorn app:app --host 0.0.0.0 --port 8100
curl -s localhost:8100/health
curl -s -X POST localhost:8100/runs \
-H 'content-type: application/json' \
-d '{"target":"octane.dev","mode":"stress",
"branches":140}'
curl -s localhost:8100/stats
curl -N localhost:8100/events
Single file, orchestrator/app.py, no database, no worker fleet. The whole engine is one module you can read in ten minutes.
08 · Paywall
Buy the licence, arm the house tonight
The free tier tells you where to look, the same way the framework always did. Everything that makes it a weapon, the offence, the loop and the read-back, sits behind the paywall. Founding pricing is locked for the life of the licence and the seat count never goes up.
Recon
- 1 target scan per day
- Passive sources only
- Findings on screen
- Offensive push
- Intelligence loop
- Read aloud, clip export
Home Shield
- Agent for the whole house, 14 surfaces
- Offensive push, attacker profiled at the source
- Intelligence loop, 40s swarm sync
- Read aloud plus 9:16 clip export
- Unlimited scans, all 2,148 sources
- Terminal install, host-bound key
Node Cluster
- Everything in Home Shield, five hosts
- Family footprint watch, kids' handles
- Vendor and supply chain sweep
- Posture report, monthly, read aloud
- Priority signature shipping
- Private loop tenancy
The gate, in one line
Free answers "where do I look". Paid answers "who came, what they knew, and why the next one will not get in". After activation the buyer runs rifm --status and reads their own blocked-probe counter.
09 · Promo
The 48 second cut
1080x1920, 30 fps, narrated, captions burned in for sound-off scrolling. The frame-by-frame analysis is public, and it is where this page's design brief came from: brighter signal colour, one accent reserved for hits, a cut every few seconds instead of every twenty four.
Spec
| Runtime | 48.7 s |
| Frame | 1080 x 1920, 30 fps |
| Encode | H.264 365 kbps, AAC 163 kbps |
| Scenes | 7 blocks, 2 hard cuts |
| Speech coverage | 59.1% |
| Download | osint_framework_read_it_for_me.mp4 |