Download Latest Version neurosploit-v4.4.2-windows-x64.zip (6.7 MB)
Home / v4.4.1
Name Modified Size InfoDownloads / Week
Parent folder
neurosploit-v4.4.1-linux-x64.tar.gz 2026-11-16 7.0 MB
neurosploit-v4.4.1-linux-arm64.tar.gz 2026-11-16 6.9 MB
neurosploit-v4.4.1-windows-x64.zip 2026-11-16 6.6 MB
README.md 2026-11-16 2.2 kB
v4.4.1 -- Operations engineering for long runs source code.tar.gz 2026-11-16 1.5 MB
v4.4.1 -- Operations engineering for long runs source code.zip 2026-11-16 2.0 MB
Totals: 6 Items   23.8 MB 0

Five new harness modules, inspired by Artex's operating model. Each is pure, deterministic and unit-tested (562 harness tests pass).

  • Cold-node digests (cold_digest) — settled exploration nodes (completed intents, facts, negatives) that stay untouched for N planner rounds are grouped and folded into short digests for the planner prompt. Goals, findings, hints and pending work are never folded. The graph itself is never mutated; a member that becomes active again invalidates its digest. LLM digest prompt plus an offline heuristic fallback.
  • Operator constraints (constraints) — "no port scan", "passive only" and similar rules are extracted from the task text into self-contained deny/allow statements (the concrete target replaces "the current target"), injected into worker prompts, and the common classes (port scan, brute force, DoS) are blocked deterministically at command level. Constraints only restrict; they never widen what scope.rs allows.
  • Single-generator intent queue (intent_queue) — only the planner (or the operator) can create intents; workers claim, complete or fail them and can only suggest() leads. Dedupe, priority claiming that respects depends_on, bounded retries, attempt refund on release.
  • Wrap-up (wrapup) — when turns or time run out, the run gets a short settlement phase with only persistence tools open and a per-role prompt, so findings, facts and negative results already identified are saved instead of lost.
  • Retest verdicts (retest) — reproduced / fixed / inconclusive computed in code. fixed requires a comparable environment, met preconditions, a completed trigger, a passing positive control and written evidence; timeouts, WAF blocks and auth failures can never produce fixed.

Pipeline wiring (post-4.4.1): operator constraints are injected into every recon/exploit prompt and enforced on recon commands; the exploration block in prompts is folded into digests (persisted as graphs/digest-state.json). The queue, wrap-up and retest modules are available in the harness crate; their drivers land in a later release.

Source: README.md, updated 2026-11-16