// the take
NeuroSploit is a Rust CLI that runs an autonomous penetration test end to end — recon, agent selection, exploitation, validation, report — using a pool of LLMs as the 'hands' while a deterministic harness drives the methodology. It's aimed at red teamers, bug bounty hunters and security-minded dev teams who want an agentic pentest tool without hand-rolling the orchestration themselves.
The core design — re-sending every claimed finding through a scope-guarded replay engine and having a second model adjudicate before it's reported — is a real attempt to fix the hallucinated-finding problem that single-shot 'ask an LLM to hack this' tools have. Validation is deterministic where it matters: per-CWE validators check actual request/response bytes instead of trusting the model's prose. It's model-agnostic (19 providers, API key or local subscription, down to fully local ollama/llamacpp), so you're not locked into one vendor or forced to send target data off-box. The 486 agent playbooks are plain markdown files, not compiled into the binary, so adding coverage for a new CWE or chain is a text edit, not a Rust change.
The safety guarantees ('stays in scope', 'no destructive actions') are enforced by agent prompting and harness logic, not hard sandboxing — a prompt-injected target response or a scope-escape bug could still trigger unintended live actions, which matters a lot for a tool whose job is autonomous exploitation. Getting the full system described in the README running is not a one-binary affair: you need LLM access (key or subscription) plus optional Neo4j, Ghidra, MobSF, Frida and trivy/grype/syft depending on mode, so the Rust binary alone is a fraction of the actual footprint. Raw credentials the agents capture during a run are kept in a 'vault' for re-testing with no detail on how that vault is encrypted or access-controlled — a real gap for a tool whose entire purpose is harvesting live secrets. It's also young in its current form: the whole Python engine and web GUI were dropped for this Rust-only rewrite, so the validator/dedup/graph logic hasn't had much real-world runway to prove its false-positive rate against actual engagements.
We read the README, the directory tree and a key source file in October 2026. Nobody pays to be listed. How we judge →