Notes from the operator.
Engineering, product, and research notes from the team building HELIX. How we plan engagements, why every finding ships with a reproducer, and what an autonomous offensive operator looks like from the inside.
Why we built a tree-search planner for offense
A pentest is a search problem with expensive moves. Here is why we reached for MCTS and UCB1 instead of a single mega-prompt, and how pruning failed branches keeps the engine from banging on the same closed door.
ProductZero false positives: why every finding ships with a reproducer
Noise is what makes teams stop reading reports. We walk through our evidence-first pipeline, runtime corroboration, the DOUBT skeptic agent, and the copy-pasteable curl that lands with every confirmed bug.
ResearchBOLA, BFLA and the bugs scanners structurally can't see
Broken object- and function-level authorization need an attacker that understands roles and intent, not signatures. Why access-control bugs slip past scanners, and how CLEARANCE and ASCENT reason about them instead.
SecurityInside the six-layer guardrail engine
Autonomy near real systems is only defensible if the constraints come first. A tour of all six layers, scan mode, scope, destructive-action blocking, budget cap, rate limiting, and the human-in-the-loop gate.
ProductAgentless by design: testing without installing in your infra
HELIX runs engagements from the outside, you point it at a target, not at your servers. What "no agents in your infrastructure" actually means for scope, trust boundaries, and how fast a team can get started.
ResearchFrom scanner alerts to confirmed exploits
A scanner says what might be wrong; an operator proves what is. We trace one hypothesis from a raw signal, through corroboration and refutation, to a triaged finding with CVSS, CWE, and remediation attached.
Get the next note in your inbox.
Occasional, technical, no fluff. Notes on building an autonomous offensive operator.