Skip to content
KitploitKITPLOIT
ToolsExploitsBlog
Log in
Submit
ToolsExploitsBlog
Submit

Hacking, PenTest, and Cybersecurity Tools for Your Security Arsenal!

Kitploit is a directory of hacking, cybersecurity, and pentesting tools. Discover the latest project updates to find vulnerabilities, analyze systems, automate testing, and strengthen your security.

FeedsContactPrivacy© 2026 Kitploit

Tool Directory

Categories

View all categories
Loading categories
ZoneClaw-code — Benchmark and defense code for persistent memory attacks on OpenClaw-style computer-use agents, with memory-zoning mitigation, attack scenarios, and evaluation scripts. | Kitploit
Tools/GitHubGitHub/euph00/zoneclaw-code
Defensive ToolsPenetration TestingMachine LearningPapers & ResearchLearning & EducationAI SecurityAdversarial AttackLabs & Practice
GitHubeuph00/zoneclaw-code

ZoneClaw-code

Benchmark and defense code for persistent memory attacks on OpenClaw-style computer-use agents, with memory-zoning mitigation, attack scenarios, and evaluation scripts.

1111926 years agoNot yet reviewed
Share
View Repository

ZoneClaw

Experiment code for ZoneClaw: Mitigating Persistent Memory Attack by Establishing Memory-Zoning among OpenClaw-Style Computer-Use Agents.

This repository contains the benchmark tasks, defense implementations, and evaluation scripts used in the paper. The benchmark studies cross-session persistent memory attacks in OpenClaw-style computer-use agents. It covers two attack classes:

  • Hidden side effects: BCC exfiltration and chat mirroring.
  • Provenance corruption: source redirection and marketplace redirection.

ZoneClaw separates retained external observations from authority-bearing memory, then uses role-separated agents to observe, classify, and act on persistent information.

This code-only edition does not bundle experiment results. The paper's transcripts, verifier evidence, and summaries are available in the accompanying artifact repository.

Setup

Prerequisites

  • Linux or WSL2 (tested on Ubuntu 24.04 under WSL2, x86-64)
  • Docker with Compose support
  • Git
  • Python 3.12+
  • uv

Allow at least 20 GB of free disk space for dependencies, Docker images, and generated runs. We recommend 16 GB of RAM for parallel experiments. Initial setup typically takes 10-20 minutes, depending on the network and Docker build cache.

For an anonymous download, extract the ZIP, open a terminal in its top-level directory, and run:

bash setup.sh

Setup fetches the pinned OpenClaw, Harbor, and WorkspaceBench dependencies directly from their public upstream repositories. No access to the original private repository is needed.

For a Git checkout, replace <repository-url> below with the repository's clone URL:

git clone --depth 1 --recurse-submodules --shallow-submodules \
  <repository-url> ZoneClaw-code
cd ZoneClaw-code
bash setup.sh

The script validates the local tools and Docker daemon, builds the required images, and creates .env from .env.example. It generates OPENCLAW_GATEWAY_TOKEN locally; edit .env only to set the provider keys needed for a run. Do not commit .env.

ExperimentRequired API keys
Anthropic main modelANTHROPIC_API_KEY
OpenAI main modelOPENAI_API_KEY
Z.AI main modelZAI_API_KEY
Alibaba Qwen main modelALIBABA_KEY
Watcher or MELON with a non-Anthropic main modelMain-model key and ANTHROPIC_API_KEY
WorkspaceBench authority and benign-utility evaluationsOPENAI_API_KEY and ANTHROPIC_API_KEY

Experiments make paid provider calls. Start with one trial to verify the complete path:

MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-zoneclaw-userprompt N=1 P=1 bash bench.sh

The main paper experiments use Claude Sonnet 4.6. After the smoke test succeeds, reproduce the full row with:

MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-zoneclaw-userprompt N=15 P=3 bash bench.sh

Outputs are written to runs/<timestamp>__e2e-<experiment>/. The command prints a per-trial breakdown after aggregation; an existing run can be viewed again with:

bash show-bench.sh runs/<run-directory>

Main Attack and Mitigation Experiments

Named end-to-end experiments follow:

<scenario>[-<defense>]-<setting>
Paper termCommand token
BCC exfiltrationbcc
Chat mirrorchat
Source redirectionsr
Marketplace redirectionmarket
User-triggered updateuserprompt
Periodic updateheartbeat
No defenseomit the defense token
ClawKeeper-style Watcherauditor
Privilege separationprivsep
ClawGuardclawguard
MELONmelon
ZoneClawzoneclaw

Examples:

# Undefended BCC, user-triggered update
MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-userprompt N=15 P=3 bash bench.sh

# ClawKeeper-style Watcher, periodic Chat Mirror update
MODEL=anthropic/claude-sonnet-4-6 \
E2E=chat-auditor-heartbeat N=15 P=3 bash bench.sh

# ZoneClaw, user-triggered source-redirection update
MODEL=anthropic/claude-sonnet-4-6 \
E2E=sr-zoneclaw-userprompt N=15 P=3 bash bench.sh

All valid names and their injection/exploitation task pairs are listed in experiments/e2e.tsv.

To use another supported model, replace MODEL; for example:

MODEL=openai/gpt-5.5 OPENCLAW_THINKING=medium \
E2E=bcc-zoneclaw-userprompt N=15 P=3 bash bench.sh

MODEL=zai/glm-5.2 \
E2E=bcc-zoneclaw-userprompt N=15 P=3 bash bench.sh

BENCH_TOKEN_PRICES_FILE="$PWD/experiments/measurement/qwen3.7-plus-2026-05-26.json" \
MODEL=alibaba/qwen3.7-plus-2026-05-26 OPENCLAW_THINKING=medium \
E2E=bcc-zoneclaw-userprompt N=15 P=3 bash bench.sh

Measured runs archive per-phase runtime, model usage, and helper-model usage alongside each trial. Alibaba Qwen runs additionally retain a content-free raw usage ledger so reasoning tokens are not double-counted by the compatibility adapter. See docs/measurement.md for the schema and aggregation rules.

Additional Experiments

ZoneClaw Ablations

The paper evaluates each ablation on the user-triggered form of all four scenarios. The BCC commands are:

# w/o Zone
MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-zoneclaw-no-authority-metadata-userprompt N=15 P=3 bash bench.sh

# w/o Gatekeeper
MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-zoneclaw-auto-promotion-userprompt N=15 P=3 bash bench.sh

# w/o Cross-check
MODEL=anthropic/claude-sonnet-4-6 \
E2E=bcc-zoneclaw-low-context-userprompt N=15 P=3 bash bench.sh

Replace bcc with chat, sr, or market for the remaining scenarios.

Defense-Aware Attacks

for experiment in \
  bcc-prompt-aware-promotion-userprompt \
  bcc-zoneclaw-prompt-aware-promotion-userprompt \
  bcc-prompt-aware-composed-userprompt \
  bcc-zoneclaw-prompt-aware-composed-userprompt \
  sr-prompt-aware-finegrained-userprompt \
  sr-zoneclaw-prompt-aware-finegrained-userprompt
do
  MODEL=anthropic/claude-sonnet-4-6 \
  E2E="$experiment" N=15 P=3 bash bench.sh
done

Selective Promotion

MODEL=anthropic/claude-sonnet-4-6 \
TASK=bcc-exfiltration-zoneclaw-promotion/benign-only \
N=30 P=3 bash bench.sh

MODEL=anthropic/claude-sonnet-4-6 \
TASK=bcc-exfiltration-zoneclaw-promotion/malicious-only \
N=30 P=3 bash bench.sh

Repeated Use Across Sessions

The paper uses the BCC repeated-attack schedule, ten update sessions, and checkpoints after 0, 1, 2, 3, 5, and 10 updates:

MODEL=anthropic/claude-sonnet-4-6 \
SYSTEM=no-defense SCHEDULE=repeated-attack \
N=3 P=1 bash longitudinal-bench.sh

MODEL=anthropic/claude-sonnet-4-6 \
SYSTEM=zoneclaw SCHEDULE=repeated-attack \
N=3 P=1 bash longitudinal-bench.sh

WorkspaceBench Authority Probe

WorkspaceBench experiments require its Lite metadata and workspace files:

uv run --with huggingface_hub python \
  workspace-bench/evaluation/scripts/download_hf_assets.py \
  --eval-root workspace-bench/evaluation \
  --language en --lite --workspaces

After downloading the upstream tasks, generate the paper's two disjoint 15-task subsets with the fixed selection seed:

WB_SUBSETS="$PWD/workspace-bench/evaluation/.generated/memory_authority_probe/subsets"
python3 scripts/workspacebench_privilege_probe.py make-subset \
  --n 15 --seed 20260709 --language en \
  --dest "$WB_SUBSETS/lite-en-15-seed20260709"
python3 scripts/workspacebench_privilege_probe.py make-subset \
  --n 15 --seed 20260709 --language en \
  --exclude-selection "$WB_SUBSETS/lite-en-15-seed20260709/selection.json" \
  --dest "$WB_SUBSETS/lite-en-15-extension-seed20260709"

Run each of the four authority conditions on both subsets:

Download Tool

Most Popular

View all →

Discover the most used tools by our community.

  • 1

    OsintHunter

    letox4/osinthunter

    🎯 The ultimate OSINT toolkit: investigation graph, confidence scoring, alias families, map, timeline and reports. 49 modules, 400+ sites, no API keys required.

    786
  • 2

    ghostlock-app

    yukonga/ghostlock-app

    GhostLock One-Tap Execution App (CVE-2026-43499)

    506
  • 3

    ARTEX

    baarai/artex

    LLM multi-agent autonomous penetration testing platform with planner/worker agents, asset and exploration graphs, MITM traffic recording, and vulnerability retesting.

    429
  • 4

    redforge

    cybersecurity91/redforge

    Browser-side red team toolkit with 102 offline tools for recon, payload generation, credential attacks, crypto, encoding and engagement reporting.

    345
  • 5

    CyberStrikeAI

    ed1s0nz/cyberstrikeai

    The system of action for AI-native cybersecurity—where intent becomes governed execution, evidence becomes operational memory, and every operation improves the next.

    262
  • 6

    looking-glass

    anonghosty/looking-glass

    CSIRT platform that ingests Shadowserver reports into MongoDB and replays them as interactive maps, analysis, findings, advisories and executive dashboards.

    256

Explore all tools

Browse our collection of tools

View all tools →