Defensive security · In development
Rides alongside your application.
Outrider is a swarm of specialised defensive agents that watch your application's edges, link misuse spread thin across many accounts, and propose a response. Nothing happens to a real user until a person says yes.
The problem
The attacks that hurt most stay under every threshold.
Credential stuffing, scraping, promo abuse and account takeover rarely come from one noisy account. They come from dozens of quiet ones, each staying just below the limits a rule would trip.
Rules that judge one account at a time see nothing wrong. The pattern only shows when you look across accounts: the same devices, the same networks, the same targets, in the same few minutes.
How it works
Five stages, one shared picture.
Specialised agents each watch one slice of your traffic and share evidence. Only correlated incidents reach the AI, and only a person can approve anything with real impact.
-
Detect
Identity and API agents watch sign-ins, sessions and requests, each on its own shard, with fast, deterministic rules.
-
Correlate
Correlation agents link accounts that share devices, networks, targets and timing into a single incident.
-
Investigate
A self-hosted open-weight model reviews each incident and writes an assessment. Its output is advice, never evidence.
-
Verify
Independent policy checks test every proposed action against the evidence and reject anything that does not hold up.
-
Respond
Low-impact actions run in dry-run by default. Anything that touches a real account waits for a person to approve it.
human approval
Principles
Built the way a defender should be.
An automated defender that can act on your users needs stronger guarantees than the attackers it is chasing. These are not settings. They are how Outrider is built.
A person decides
Dry-run is the default. High-impact actions need human approval, every decision is audited under the approver's verified identity, and a global emergency stop halts everything.
No AI on the hot path
Detection and correlation are deterministic, fast and explainable. The model only sees incidents that are already correlated, and it can never count as evidence on its own.
Your data stays yours
Outrider runs in your own cloud with a self-hosted open-weight model. No third-party AI API sees your users' traffic.
Logs are data, not orders
Attackers write text too. Nothing Outrider reads in traffic or logs is ever treated as an instruction, and agents have no shell or infrastructure access.
Defence only
Outrider protects systems you own or are authorised to defend. It does not scan, probe or hack back against anyone.
Locked down by default
Internal services refuse traffic from outside the private network, workloads use short-lived identities instead of keys, and the model server is never exposed publicly.
Lab results
What we have measured so far.
Measured on a synthetic application with synthetic users and attack campaigns, averaged over five seeds per profile. These are lab numbers, not production claims.
A live run on Google Cloud attributed all 30 suspicious accounts with no false positives. Coordination costs some precision: the swarm averaged 0.4–1.4 false-positive accounts per run, which is exactly why restrictive actions wait for a person.
Where we are
A working prototype, looking for its first real traffic.
- Built and tested. Every agent type, the simulator, the dashboard and the benchmark, with automated tests.
- Deployed. Running on Google Cloud with a self-hosted model on a GPU, behind a locked-down private network.
- Approvals proven. Dry-run, human approval, live actions and the emergency stop tested end to end.
- Next: shadow mode. Reading a real application's logs and proposing actions, without ever applying them.
Run Outrider in shadow mode
If you run an application that suffers from account abuse, we would like to run Outrider alongside it in read-only shadow mode. It watches and proposes, nothing is applied, and you see what it would have caught.
Email us