Evidence-backed hypothesis trees
A tree can organize alternative mechanisms and their prerequisites. It is an
analysis representation, not an execution plan or a mandated pentest lifecycle.
Root nodes name the security question; children identify necessary conditions,
competing explanations or concrete observations. Link every observed node to its
source and keep proposed tests distinct from results.
Modeling constraints
Use identifiers stable enough to retain provenance when an assumption changes.
A child is relevant only when it can change the parent conclusion. Avoid invented
services, ports or account privileges. Record contradicted nodes and why they were
refuted rather than deleting the evidence. A closed/filtered observation at one time
may invalidate a current path, not every possible path or future state. Keep uncertain,
refuted, supported and untested claims distinct.
Distilling evidence
For a network scan retain endpoint, protocol, version source, timing and filtering
limits. For HTTP retain route, method, identity, input, status, relevant content and
comparison control. For source review retain revision, location and caller-to-sink
conditions. Summaries should remove repetitive payload bytes, not evidence qualifiers.
A 403 does not prove a real protected file exists, and a shell-like prompt does not
prove access without an attributed execution observable.
Comparing branches
Choose discriminating questions by information gain, available evidence and task
constraints. Similar endpoint names do not establish a shared root cause; a credible
cause can motivate bounded variant analysis. Never derive authorization from a
branch's rank. The runtime's objective graph remains authoritative for execution;
workflow stage order, readiness, dispatch and reporting do not belong in this tree.
1---2name: pentest-task-tree3description: "Iterative PTT (Penetration Testing Tree) session reasoning — model evidence, alternative mechanisms, and prerequisites without prescribing an engagement lifecycle."4allowed-tools: Read Write5metadata:6 source: https://skillmd.com/skills/purpleailab/pentest-task-tree7 skillogy:8 version: 19 subdomain: orchestration10 when_to_use: "PTT, task tree, what to do next, next task, session reasoning, iterative pentest, live pentest session, update task list, pentest tree, task selection"11 tags: ptt, task-tree, session-reasoning, iterative, decision-making, llm-pentest, pentestgpt12 upstream_ref: "PentestGPT (Gelei Deng et al., USENIX Security 2024) — pentest task tree methodology spanning the full ATT&CK Enterprise kill chain"13---1415# Evidence-backed hypothesis trees1617A tree can organize alternative mechanisms and their prerequisites. It is an18analysis representation, not an execution plan or a mandated pentest lifecycle.19Root nodes name the security question; children identify necessary conditions,20competing explanations or concrete observations. Link every observed node to its21source and keep proposed tests distinct from results.2223## Modeling constraints2425Use identifiers stable enough to retain provenance when an assumption changes.26A child is relevant only when it can change the parent conclusion. Avoid invented27services, ports or account privileges. Record contradicted nodes and why they were28refuted rather than deleting the evidence. A closed/filtered observation at one time29may invalidate a current path, not every possible path or future state. Keep uncertain,30refuted, supported and untested claims distinct.3132## Distilling evidence3334For a network scan retain endpoint, protocol, version source, timing and filtering35limits. For HTTP retain route, method, identity, input, status, relevant content and36comparison control. For source review retain revision, location and caller-to-sink37conditions. Summaries should remove repetitive payload bytes, not evidence qualifiers.38A 403 does not prove a real protected file exists, and a shell-like prompt does not39prove access without an attributed execution observable.4041## Comparing branches4243Choose discriminating questions by information gain, available evidence and task44constraints. Similar endpoint names do not establish a shared root cause; a credible45cause can motivate bounded variant analysis. Never derive authorization from a46branch's rank. The runtime's objective graph remains authoritative for execution;47workflow stage order, readiness, dispatch and reporting do not belong in this tree.