sthrip is a research lab for autonomous offense.
We train our own models with reinforcement learning and spawn giant swarms of agents that break web apps like a red team — planning, probing, and chaining attacks end to end, gated to your scope.
Nothing reaches the report until an independent critic and a validator both agree — and the validator mechanically reproduces the finding. We are not a scanner. We ship verified evidence.
Agent depth benchmark
17 planted vulnerability classes behind a deterministic fixture net. The swarm found every one — zero false positives on decoys.
| # | Metric | Result |
|---|---|---|
| 1 | Recall — planted classes100% | 100% |
| 2 | Precision — live decoys100% | 100% |
| 3 | Recall — baseline planner5.9% | 5.9% |
RL attack-graph
Our own RL-trained planner ranks every attack path by expected value. Failed techniques get discounted; confirmed ones escalate.
Get access →Giant swarms
Up to 400 parallel agents per engagement, with per-host throttling that backs off the moment a target pushes back.
Get access →Lab ladder
A five-tier adversarial lab (L0–L4) measures the swarm before it earns autonomy. Every violation ratchets it back down.
Get access →Watch a giant swarm run a live engagement.
A scoped campaign against our own lab — from the RL-planned first probe to mechanically reproduced findings.