Handcuff — an antivirus for AI agents

Handcuff is a security-first observability and enforcement tool for AI agents. It records what an agent does at the OS level in a tamper-evident, hash-chained log — and it can stop malicious content before an agent ever sees it, through a detonate-and-verdict sandbox and, on Windows, a signed kernel driver that blocks file opens and process launches in Ring-0.

Full architecture diagrams, TUI screenshots, and the complete guide live on the Handcuff GitHub repository — this page covers install and the essentials.

Key features

• Tamper-evident recording — every event is SHA-256 hash-chained and the session head is Ed25519-signed. handcuff verify re-validates the chain from the live database or an exported report.
• Sandbox analysis before admission — SandboxAnalyzer gives every intercepted download (and MCP tool payload) an approve/reject verdict before the agent receives it: static risk-indicator analysis always, plus detonation in a hardened, network-less Docker container when available. Falls back to static-only with no Docker installed — nothing breaks.
• Kernel-level enforcement (Windows, optional) — a signed minifilter driver denies file opens and process launches for anything flagged, even for a process that never calls into Handcuff. Content-hash matching catches a copy/rename of previously-flagged bytes. The deny-list persists and re-applies itself across reboots.
• MCP proxy firewall — handcuff mcp intercepts tools/call requests, scans them with a regex injection detector plus an optional local LLM judge (Ollama, 100% local), and returns a proper JSON-RPC error for a blocked call instead of forwarding it.
• Prompt-injection circuit breaker — handcuff.gateway.Gateway implements taint-tracking: once untrusted content is read, further network egress and out-of-workspace reads are denied for a window, cutting Simon Willison's "lethal trifecta" before it completes.
• Live dashboard — handcuff tui shows the event stream, risk score, alerts, and a kernel-enforcement panel in real time.
• Local-only — nothing leaves your machine except an optional, content-light webhook you configure yourself.

Install

Requires Python 3.10+.

pip install handcuff

Optional extras, install only what you need:

pip install "handcuff[proxy]"      # mitmproxy, for the download-interception proxy
pip install "handcuff[sandbox]"    # docker SDK, for dynamic sandbox detonation
pip install "handcuff[mcp]"        # mcp package
pip install "handcuff[langchain]"  # langchain-core, for the HandcuffHandler callback

Quickstart

handcuff init

Environment readiness:
[ ok ] Config file: ~/.config/handcuff/config.yaml
[ -- ] mitmproxy (download proxy): mitmdump not found
[ -- ] Docker (dynamic sandbox): not available
[ -- ] Ollama (LLM judge): not reachable
[ -- ] Kernel driver (Ring-0 enforcement): not loaded

Each unchecked item is optional - the tool works with whatever is available.

handcuff watch -- python my_agent.py

session 01KXG26CAXY0XB65RMG05EGHFD closed

handcuff verify 01KXG26CAXY0XB65RMG05EGHFD

chain: OK
signature: OK

Check a file before you trust it, no Docker required:

handcuff sandbox analyze ./suspicious_script.sh

REJECTED - high-risk static indicators
method: static-fallback
indicators: disk_wipe

The CLI, at a glance

handcuff watch -- <command>          # record OS-level activity for <command>
handcuff mcp -- <command>            # MCP stdio proxy with tool-call scanning
handcuff proxy                       # download interception proxy
handcuff sandbox analyze <file>      # detonate/analyze verdict for a file
handcuff kernel status               # kernel driver load state + quarantine count
handcuff kernel quarantine <path>    # deny file opens for <path>
handcuff kernel watch <dir>          # background auto-quarantine
handcuff tui                         # live dashboard
handcuff verify <session_id>         # re-validate hash chain + signature
handcuff doctor                      # full environment diagnostics

Every command supports --help.

Python integration

from handcuff.integrations.langchain import HandcuffHandler

handler = HandcuffHandler()
result = my_agent.invoke(
    {"input": "Do some research"},
    config={"callbacks": [handler]},
)
handler.close()

from handcuff.sandbox.manager import DockerSandbox

# Hardened defaults: no network, read-only root fs, capability drop,
# memory/pids/CPU ceilings.
with DockerSandbox(allow_network=True) as sandbox:
    sandbox.execute("curl -sO https://example.com/untrusted_data.csv")
    # Runs the injection scanner + a native-executable check before the
    # file ever reaches the host; raises QuarantineViolation and discards
    # it on a match instead.
    sandbox.copy_quarantined(
        container_path="/sandbox/untrusted_data.csv",
        host_path="./host_workspace/clean_data.csv",
    )

Honest scope

• The regex/heuristic injection scanner is a tripwire against known patterns, not a semantic guarantee.
• The Gateway and MCP proxy only secure agents whose tool-calling layer routes through them — an agent with raw shell access bypasses both. The kernel driver is the only layer that doesn't depend on agent cooperation, and it requires Windows plus a one-time signed build; nothing else in Handcuff requires it.
• Dynamic sandbox execution runs scripts, not opaque binaries, and container isolation is a strong mitigation, not a zero-risk guarantee.

Full documentation

Architecture diagrams, the complete CLI/config reference, the kernel driver build/sign workflow, and contributing guidelines: github.com/Mister2005/Handcuff.

Handcuff currently prioritizes the Windows ecosystem (psutil, watchdog, Windows ACL/Firewall, the kernel driver). macOS and Linux are partially supported by the underlying capture abstractions, but enforcement (ACL/firewall/kernel) is Windows-only today.