NeuroSploit Files

NeuroSploit v4.4.0

Recon-Driven Agent Selection

• Negative filtering: agents for technologies NOT detected in recon are penalized (-10 score) instead of running blindly
• Stack-aware scoring: WAF/CDN detected → boost bypass/smuggling agents (+8), reduce basic injection baseline
• 30+ surface detection flags: web, login, API, GraphQL, JWT, OAuth, WAF, CDN, SPA, PHP, Java, .NET, Python, Node, cloud, k8s, docker, DB, AI, CORS, template, cache, XML
• 60+ signal mappings (was ~23): covers frameworks, CDNs, WAFs, patterns
• Smarter LLM selector prompt: explicit elimination rules, hard exclusions for irrelevant agents, 6000 char recon context (was 3000)

MCP Hot-Reload

• Watch mcp.servers.json for changes mid-run — add/remove MCP servers without stopping the engagement
• Each agent gets fresh CLI subprocess reading .mcp.json from disk
• REPL commands: add_server() / remove_server() for runtime MCP management

Reference Prompts & Autonomy Levels

• reference_prompts: inject methodology files (.md) into every agent prompt as supplementary testing guidance
• autonomy levels: cautious, standard, aggressive, full — controls tool installs, creative attacks, and guardrail strictness

Findings Preservation

• Vote-rejected findings are now KEPT with review_status: "rejected" instead of being silently discarded
• All candidate findings visible in the web console with clear status indicators
• rejected.jsonl still written for backward compatibility

Web Console

• Working search/filter on Pentests list (by ID, target, status)
• Working search/filter on Issues page (text search + severity dropdown + status dropdown)
• Version v4.4.0

Execution Resilience

• Error classification with exponential backoff
• Crash salvage for interrupted runs
• LLM dedup, budget warnings, agent nudge, image budget, and run resume