ZeroForge

An adaptive binary-0day research brain that sits on top of HexStrike AI.

HexStrike is an excellent executor: 150+ security tools wrapped behind an HTTP API and MCP. But it is stateless — every tool returns raw text to the LLM and the target model is forgotten the moment the call returns, and its "decision engine" is a hard-coded effectiveness table that never adapts.

ZeroForge adds the part that was missing: memory and reasoning.

   ┌── LLM (Claude via MCP) ──┐
   │   reasons & decides       │
   └────────────┬──────────────┘
                │  small stateful verbs (zf_ingest / zf_plan / zf_assess …)
        ┌───────▼─────────┐
        │   ZeroForge     │   persistent knowledge graph  (SQLite)
        │   • model       │   adaptive planner (state → ranked actions)
        │   • planner     │   action log (never repeat work)
        └───────┬─────────┘
                │  executes via …
     ┌──────────▼───────────┐        ┌───────────────┐
     │  HexStrike /api/tools │  or →  │ local tools   │  (fallback if HS down)
     └──────────────────────┘        └───────────────┘

Why it's more powerful than raw HexStrike

	HexStrike	ZeroForge
Target model	none — raw tool text per call	persistent knowledge graph that accumulates across runs
Tool selection	static effectiveness table	adaptive planner reacting to discovered facts
Repeated work	re-runs freely	action log dedup — never repeats a completed step
Reasoning	in the LLM only, each turn cold	LLM + a growing state it can query
Backend	its own wrappers	reuses HexStrike, with local fallback

Install

python3 -m pip install requests            # only hard dep for the core
# MCP mode additionally needs fastmcp (already in the HexStrike venv)

Use — CLI

python3 -m zeroforge.cli ingest ./target      # characterize + enumerate + gadgets
python3 -m zeroforge.cli assess               # narrative state
python3 -m zeroforge.cli plan                 # ranked next actions + rationale
python3 -m zeroforge.cli exploit ./target     # autonomous ret2win -> real shell
python3 -m zeroforge.cli show sink            # dump graph nodes
python3 -m zeroforge.cli log                  # what has already run

Autonomous exploitation (proven end-to-end)

Against a no-canary stack overflow with a win()/system("/bin/sh") function, zeroforge exploit runs the whole chain by itself — no offsets hard-coded:

[*] Attempting autonomous ret2win on ./vuln_test
    detail: ret2win -> win@0x401166 offset=72 align_ret=0x401016
[+] EXPLOIT SUCCESSFUL — code execution proven:
      | ZEROFORGE_PWNED_1000
      | uid=1000(benny) gid=1000(benny) groups=1000(benny),...
[+] Reproducible exploit written to: ./vuln_test.zeroforge_exploit.py

It (1) detects the input vector from the discovered sinks, (2) computes the saved-RIP offset by crashing the target under GDB with a cyclic pattern, (3) locates a win/system function, (4) builds a stack-aligned payload and (5) runs the binary to obtain a shell — then records GOAL ACHIEVED: shell into the knowledge graph and emits a standalone pwntools script. tests/test_exploit.py compiles a fresh target and asserts the shell.

Example against a scanf/strcpy/system("/bin/sh") binary with no canary:

[ 98] Trace reachability to system()/popen() (exploit)
       ↳ system()/popen() present and a /bin/sh string exists — likely a one-shot command-injection path.
[ 93] Dynamically confirm stack overflow (offset) (exploit)
       ↳ Unbounded copy sink present. No stack canary → saved-return overwrite is viable.
[ 82] Probe format-string primitive (exploit)
       ↳ Non-constant format sink present → test for arbitrary read/write & leaks.

Use — MCP (drive it from Claude)

Point your MCP client at zeroforge.mcp_server. Verbs: zf_ingest, zf_plan, zf_assess, zf_show, zf_note, zf_run_hexstrike.

{
  "mcpServers": {
    "zeroforge": {
      "command": "/home/you/hexstrike-ai/hexstrike-env/bin/python",
      "args": ["-m", "zeroforge.mcp_server"],
      "env": { "ZEROFORGE_DB": "engagement.db", "HEXSTRIKE_URL": "http://localhost:8888" }
    }
  }
}

Status

v0.2 — working core: knowledge graph, triage capabilities (file/checksec/sinks/ strings/ROP gadgets), adaptive planner, CLI, MCP server (7 verbs incl. zf_exploit), and autonomous ret2win exploitation proven end-to-end. Roadmap: ret2libc / full ROP chaining, format-string primitive, taint-based reachability (angr), fuzzing orchestration + crash triage, agentic run-loop.

Legal

For authorized security research, CTFs and pentesting engagements only. You are responsible for having permission to test any target.