Raze
Offensive-security triage agent giving typed judgments on exploitability, impact and next attack path, gated by validators.
Overview
Raze is a reasoning and triage layer for authorized offensive-security work such as pentests, red-team engagements, bug-bounty triage and CTFs. It has two parts: a System One-style model interface (Raze) that maps target and engagement state to a typed judgment with a probability, and an agent (RazeAgent) that orchestrates the model, deterministic analyzers and validators behind an authorization boundary. It asks questions such as whether a finding is exploitable, what its real impact is, which attack path comes next, and whether it is a duplicate, known or theoretical issue.
Design principles are typed output over prose, reported and measured calibration (ECE), deterministic validators that can downgrade a finding regardless of probability, and human-in-the-loop by default with no unattended exploitation. A heuristic backend needs no model and, on the bundled 55-case dataset, the README reports accuracy 0.82, precision 1.00 and recall 0.67; a Claude backend is available, and a TypeSafe System One call or local model can plug in. The README calls it an early scaffold.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this finding exploitable? | — |
| Score | What is the real impact of this finding? | — |
| Choice | Which attack path should be tried next? | — |
| Choice | Is this a duplicate, known or theoretical issue? | duplicate · known · theoretical |
How it works
- Load findings and engagement state within the authorised scope
- Ask the backend (heuristic, Claude, or System One) for typed judgments
- Deterministic validators downgrade anything they can disprove
- Rank and propose next steps for the operator to approve
Strengths
- Typed, schema-validated judgments
- Heuristic backend works without any model
- Validators and human approval gate all output
Caveats
- Early scaffold per the README
- Jev itself is not the default backend
- Benchmark is a small 55-case dataset
Getting started
pip install -e . && raze plan benchmarks/offsec_dataset.json --backend heuristic --top 10At a glance
- Type
- Project
- Status
- prototype
- Language
- Python
- Runtime
- Library, CLI and agent skill
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Offensive-security agent built on Raze, a System One model in the Jev family — turns target state into typed, calibrated judgments (exploitability, impact, novelty, next action) gated by deterministic validators.
hellogumbo/awesome-jev
Frequently asked questions
What does Raze decide?
Exploitability, impact, novelty and next attack path for findings, as typed judgments with probabilities.
Is it an autonomous attack tool?
No; the README says it proposes and the operator disposes, for authorized work only.
Does it need Jev?
No; it runs on heuristics or Claude, and a TypeSafe System One call can be plugged in.
Related projects
Unix filter for security alert triage: JSONL alerts in, Jev answers five typed questions, code routes each alert.
ProjectbetaRustEvent-stream layer labelling signups, logins and checkouts as human, AI agent, crawler or bot, with a bad-intent probability.
AppprototypeJavaScriptIncident-scoping tool that starts from one malicious process and asks Jev which linked telemetry belongs to the same incident.
DemoprototypeSREGym blog post on Kubernetes incident diagnosis with no LLM agent: Jev picks a root cause from collected evidence.
ArticleresearchGo CLI for managing WordPress sites at scale, with typesafe commands that ask Jev questions and a scan --triage that ranks malware findings.
ProjectproductionGoCisco's security scanner for agent skills, with a System One analyzer as an advisory tier that cannot emit findings or change severity.
Agent skillproductionPython