ailock 🔐

Cryptographic lockfile for AI/LLM Python dependencies. Like package-lock.json, but for litellm, langchain, openai, transformers, and friends.

---

🚨 Active Threat: TeamPCP Campaign (March 2026)

The supply chain campaign is active RIGHT NOW. Three ecosystems compromised in 8 days:

Date	Package	Versions	CVE	SHA256 (malicious wheel)
Mar 19	trivy-action (GitHub)	75 of 77 tags	CVE-2026-33634	—
Mar 24	litellm	1.82.7, 1.82.8	CVE-2026-33634	8395c326... / d2a0d5f5...
Mar 27	telnyx	4.87.1, 4.87.2	CVE-2026-33634	7321caa3... / cd081158...

ailock audit has all four malicious package hashes pre-loaded. Run it now.

---

The Attack

On March 24, 2026 at 10:39 UTC, two versions of LiteLLM were published to PyPI by TeamPCP after they compromised maintainer credentials via a poisoned Trivy CI scanner.

litellm 1.82.7 — malicious payload embedded in proxy_server.py. Triggers on import. Three-stage attack: credential harvest (SSH keys, .env files, AWS tokens, Kubernetes secrets) → lateral movement → persistent systemd backdoor. Exfiltrates to models.litellm.cloud as tpcp.tar.gz.

litellm 1.82.8 — even worse: adds litellm_init.pth to site-packages. This fires on every Python startup — including pip, IDE language servers, anything. No import required.

On March 27, the stolen credentials from LiteLLM victims were used to compromise telnyx (4.87.1, 4.87.2). WAV steganography delivers platform-specific payloads from 83.142.209.203:8080.

The version number matched. The package name matched. Standard pip install --require-hashes passed — because the attacker used legitimate credentials to publish malicious bytes. Only a pre-attack lockfile would have shown the hash change.

ailock verify would have caught it in CI, before it hit prod.

---

Install

pip install ailock

Quick Start

# 1. Generate a lockfile for your AI/LLM dependencies
ailock generate

# 2. Commit it
git add .ailock && git commit -m "chore: add ailock lockfile"

# 3. Verify in CI — fails loudly if ANY hash changed
ailock verify

# 4. Check against known-bad versions database
ailock audit

---

How It Works

ailock generate

Scans your requirements.txt / pyproject.toml, identifies AI/LLM ecosystem packages, fetches the SHA256 hashes from PyPI's JSON API for each pinned version, and writes a .ailock lockfile:

{
  "schema_version": "1",
  "generated_at": "2024-11-15T09:41:22Z",
  "packages": {
    "litellm": {
      "version": "1.82.6",
      "hashes": [
        "sha256:a7c3f8e2b1d4..."
      ],
      "source": "pypi"
    },
    "langchain-core": {
      "version": "0.3.1",
      "hashes": [
        "sha256:f4e9a1b5c2d3..."
      ],
      "source": "pypi"
    }
  }
}

ailock verify

Re-fetches hashes from PyPI and compares. If a package's hash changed since you ran generate, that package was modified after it was released.

✓  All 12 packages verified — integrity intact

🚨 SUPPLY CHAIN ATTACK DETECTED

1 package has been modified since your lockfile was generated.
DO NOT deploy. Investigate immediately.

┌─────────┬─────────┬────────────────────────────────────────────────────┐
│ Package │ Version │ Issue                                              │
├─────────┼─────────┼────────────────────────────────────────────────────┤
│ litellm │ 1.82.7  │ 2 new hashes added after lockfile was generated   │
└─────────┴─────────┴────────────────────────────────────────────────────┘

ailock audit

Cross-checks your locked packages against the community known-bad.json database. Ships with the current TeamPCP campaign entries plus historical attacks:

Package	Version	Severity	CVE	Attack
litellm	1.82.7	CRITICAL	CVE-2026-33634	Credential harvester + Kubernetes lateral movement
litellm	1.82.8	CRITICAL	CVE-2026-33634	.pth persistence — fires without any import
telnyx	4.87.1	CRITICAL	CVE-2026-33634	WAV steganography payload delivery
telnyx	4.87.2	CRITICAL	CVE-2026-33634	WAV steganography payload delivery
ultralytics	8.3.41	HIGH	—	CI/CD pipeline cryptominer

Real SHA256 hashes for all malicious wheels are included. The database is fetched fresh from GitHub on each run. Community PRs welcome — known-bad.json.

---

GitHub Actions — One-Liner

- name: Verify AI/LLM dependency integrity
  run: pip install ailock && ailock verify && ailock audit

Full workflow:

name: Supply Chain Security

on: [push, pull_request]

jobs:
  ailock:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4

      - name: Set up Python
        uses: actions/setup-python@v5
        with:
          python-version: "3.11"

      - name: Install dependencies
        run: pip install -r requirements.txt

      - name: Install ailock
        run: pip install ailock

      - name: Generate lockfile (first run)
        # Only needed if .ailock doesn't exist yet
        run: |
          if [ ! -f .ailock ]; then
            ailock generate
            echo "Generated .ailock — commit this file to your repo"
            exit 1
          fi

      - name: Verify package integrity
        run: ailock verify

      - name: Audit against known-bad database
        run: ailock audit

---

What ailock protects against

✅ Post-release replacement attacks (the LiteLLM scenario)

Package published clean → attacker gains PyPI access → replaces the wheel with malicious code under the same version number.

ailock verify catches this because the SHA256 changes.

✅ Known compromised versions

ailock audit cross-checks against a curated database of confirmed malicious versions. Catches packages you may have locked before the attack was public.

✅ Typosquats with pinned versions

If you accidentally pinned langchian==0.1.0 instead of langchain, the lockfile records it — and ailock audit flags known typosquats.

❌ What it doesn't protect against

• Packages you haven't locked (unpin them at your peril)
• Brand new attacks not yet in the known-bad database
• Compromised packages where the attacker maintains the same hash (very hard but possible via pre-image attacks — SHA256 makes this computationally infeasible today)

---

Supported Package Ecosystem

ailock automatically recognises 100+ AI/LLM packages including:

LLM SDKs: litellm, openai, anthropic, cohere, mistralai, groq, replicate, together LangChain: langchain, langchain-core, langchain-community, langgraph, langsmith LlamaIndex: llama-index, llama-index-core, llama-index-llms-* HuggingFace: transformers, huggingface-hub, tokenizers, accelerate, peft, datasets Vector Stores: chromadb, pinecone-client, weaviate-client, qdrant-client Frameworks: haystack-ai, semantic-kernel, autogen, crewai, dspy Utilities: tiktoken, sentence-transformers, instructor, outlines, guidance

Use --all-packages to lock your entire dependency tree.

---

CLI Reference

Usage: ailock [OPTIONS] COMMAND [ARGS]...

  ailock — cryptographic lockfile for AI/LLM dependencies.

Commands:
  generate  Scan AI/LLM dependencies and write a cryptographic lockfile.
  verify    Verify installed packages against .ailock cryptographic hashes.
  audit     Cross-check packages against known-bad versions database.

ailock generate

Options:
  -o, --output PATH     Output lockfile path.  [default: .ailock]
  -a, --all-packages    Include ALL packages (not just AI/LLM).
  -i, --from-installed  Use currently installed package versions.
  -f, --file PATH       Specific requirements file(s) to scan.
  --no-hashes           Skip PyPI hash fetching (faster but less secure).

ailock verify

Options:
  -l, --lockfile PATH   Path to .ailock lockfile.  [default: .ailock]
  --fail-on-missing     Exit with error if any package not on PyPI.
  --json-output         Output results as JSON (for CI integration).

ailock audit

Options:
  -l, --lockfile PATH   Path to .ailock lockfile.  [default: .ailock]
  --offline             Use only bundled database (no network call).
  --db-url TEXT         URL to known-bad JSON database.
  --json-output         Output results as JSON (for CI integration).
  --show-db             Print the full known-bad database and exit.

---

Contributing to the Known-Bad Database

The ailock/data/known-bad.json file is community-maintained at github.com/midnightrunai/ailock.

If you discover a compromised package version:

1. Fork github.com/midnightrunai/ailock
2. Add an entry to ailock/data/known-bad.json
3. Include: package name, version, SHA256 hashes (if known), description, references
4. Open a PR

Format:

{
  "package": "some-ai-package",
  "version": "1.2.3",
  "hashes": ["sha256:..."],
  "severity": "CRITICAL",
  "cve": "CVE-2024-XXXXX",
  "description": "What happened and why it's bad.",
  "reported_at": "2024-11",
  "references": ["https://..."]
}

---

Why not just use pip's built-in hash checking?

pip install --hash works but requires you to:

1. Manually specify hashes in requirements.txt
2. Include hashes for every package (including transitive deps)
3. Re-generate everything on every update

ailock focuses specifically on the AI/LLM ecosystem, adds a curated known-bad database, and makes the workflow CI-native.

---

License

MIT — see LICENSE

---

Built by Midnight Run — the AI that builds while you sleep.