Foxglove

Foxglove generates high-fidelity decoy documents — honeyfiles — and optionally mints and embeds Thinkst Canary callbacks — honeytokens — into them. It pairs Anthropic's Claude (for context-aware content) with Thinkst Canary (for verifiable detonations), so defenders can drop authentic-looking bait into their environments and get real alerts the moment an adversary touches it.

Inspired by Cliff Stoll's The Cuckoo's Egg, Foxglove aligns with the MITRE Engage framework to Expose, Affect, and Elicit adversary behaviour through deception.

Note

A name with history. Foxglove was previously called Antlion — both names capture the same idea (a lure that punishes intrusion), through different metaphors. The project was renamed on 26 June 2026 to align with its "poisonous flower" theme: Digitalis purpurea is striking to look at and toxic to anyone who handles it carelessly. Bait that looks innocent, then bites.

Important

Foxglove is for authorised adversary engagement, defensive research, and authorised penetration testing only. Do not deploy deceptive artifacts on networks where you don't have explicit permission.

Glossary

• honeyfile — a decoy document produced by foxglove generate (PDF, DOCX, XLSX, etc.).
• honeytoken — a callback URL/hostname minted on a Thinkst Canary console via foxglove thinkst create. Honeytokens can be passed to foxglove generate --honeytokens so the LLM weaves them naturally into honeyfile content.

Prerequisites

• Python 3.14+
• uv for dependency management
• Anthropic API key (required for generate)
• Thinkst Canary API key (optional — only required for thinkst subcommands; see API role below)

Installation

git clone https://github.com/lopes/foxglove
cd foxglove
uv sync

Environment Setup

# Required for foxglove generate
export ANTHROPIC_API_KEY="sk-ant-..."

# Optional — only for foxglove thinkst subcommands
export THINKST_API_KEY="..."

Both follow the standard <VENDOR>_API_KEY pattern. Most operators already have ANTHROPIC_API_KEY exported for other Claude tools — the same value works here.

Which Thinkst API role?

Foxglove needs an Admin key (create/delete canarytokens are full API operations per Thinkst's role matrix). Prefer flock-scoped Admin over console-scoped — smaller blast radius if the key ever leaks. See Thinkst — How does the API work? and Flock API Keys.

Supported Formats

PDF, Encrypted PDF (password-protected), DOCX, XLSX, PPTX, Markdown, CONF, JSON, XML

Usage

The CLI has two subcommands:

foxglove generate ...    # produce a directory of honeyfiles
foxglove thinkst ...     # mint / list / remove honeytokens on a Canary console

Use case 1 — Generate a honeyfile directory (Anthropic only)

Quickest path to value. Drop a credible-looking IT documentation tree into /tmp/op_alpha.

uv run python -m foxglove generate \
  --base-dir /tmp \
  --operation op_alpha \
  --num-files 20 \
  --formats pdf,docx,xlsx,md,conf \
  --teams it,infra,financial \
  --company "mid-sized fintech in Luxembourg, ~50 employees" \
  --file-content "internal IT and finance docs: stack details (MS-AD, Zabbix, Cisco ISE, Jamf), quarterly reports, access policies"

Result: 20 honeyfiles under /tmp/op_alpha/, plus a MANIFEST.json recording the campaign parameters, the LLM-generated authors pool, and the file list. Each document carries plausible metadata (author drawn from the pool, created/modified timestamps jittered 1–25 days apart, filesystem mtime aligned).

Use case 2 — Preview a campaign before committing (dry run)

For larger operations, preview the LLM's plan without writing anything to disk or running content generation.

uv run python -m foxglove generate \
  --base-dir /tmp \
  --operation op_preview \
  --num-files 100 \
  --formats pdf,epdf,docx,xlsx,pptx,md \
  --teams legal,hr,finance,engineering \
  --company "global SaaS, ~2000 employees, Series D" \
  --file-content "corporate documents covering legal contracts, HR policies, payroll, and engineering runbooks" \
  --dry-run

The dry-run prints every planned file (path, format, summary) so you can sanity-check coverage and tone before paying for the full content-generation pass. Pair with --resume later to continue from where you left off if interrupted.

Use case 3 — The twofold workflow (honeytoken → honeyfile)

The marquee use case. Mint a Thinkst callback, then weave it into the generated content so the LLM places it as a plausible link, contact reference, or footer URL. When the adversary opens the document and follows the link, your Canary console alerts.

# 1) Mint a callback on your Canary console
uv run python -m foxglove thinkst create \
  --domain-hash <your-domain-hash> \
  --kind dns \
  --flock flock:default \
  --memo "decoy quarterly report deployment"
#  → prints the canarytoken id + callback hostname

# 2) Generate honeyfiles, embedding the callback
uv run python -m foxglove generate \
  --base-dir /tmp \
  --operation q3_decoys \
  --num-files 5 \
  --formats pdf,docx,xlsx \
  --teams finance,executive \
  --company "..." \
  --file-content "Q3 board pack, revenue breakdown, strategy memos" \
  --honeytokens <callback-hostname-from-step-1>

# 3) (later) audit foxglove-tagged honeytokens on the console
uv run python -m foxglove thinkst list --domain-hash <your-domain-hash>

# 4) (end of engagement) clean up
uv run python -m foxglove thinkst remove --domain-hash <your-domain-hash> <canarytoken-id>

Tokens minted by Foxglove are tagged in their memo with #foxglove so thinkst list filters cleanly to just the ones it created. Pass --json to thinkst list if you want machine-readable output for scripting cleanup.

How it Works

1. Context Ingestion — operator-defined parameters (target environment, teams, content theme) flow into a Pydantic-validated campaign object.
2. Campaign Planning — Claude produces a structured roadmap of file paths, formats, and summaries plus a small pool of culturally-coherent author names. Pydantic-enforced JSON guarantees the LLM output maps cleanly into internal types.
3. Content Generation — Foxglove iterates through the plan, asking Claude to write believable per-file content. When --honeytokens is set, the LLM is instructed to weave them naturally into the text — never as an out-of-band block.
4. File Assembly — the writers (python-docx, openpyxl, python-pptx, reportlab, pypdf) materialise each format with realistic document metadata (author, created/modified timestamps jittered like real authoring lifecycles) and the filesystem mtime aligned to match.

For full architectural detail — exit codes, manifest schema, resume/dry-run semantics, Canary endpoint mapping, supported honeytoken kinds — see DESIGN.md.

Contributing

See CONTRIBUTING.md. Short version: strict TDD, scoped commits, immutable data, no comments unless the why is non-obvious.

Security

For vulnerability reports and disclosure practices, see SECURITY.md.

License

MIT — see LICENSE.