AION

Scan Python repos. Open verified security fixes. Nothing you didn't ask for.

AION scans a Python repository for a focused set of high-confidence security issues, generates a deterministic patch for each one, verifies the patch in isolation (syntax + an AST assertion that the fix actually holds + an optional Semgrep re-scan), and opens a pull request only for fixes that pass. If a fix can't be proven safe, AION reports it for human review instead of touching your code.

It is intentionally small. The goal is a tool you can trust to run unattended on your repository, not a platform you have to operate.

What it actually does

• Context-aware scanning — profiles the repository (ORM, HTTP client, auth decorators, DB patterns) so findings are grounded in how your code is written, with optional LLM explanation.
• Deterministic, verified auto-fixes for these issue types:

  Issue	Fix
  hardcoded_secret	move the literal to os.getenv(...)
  raw_sqlite_query	parameterize the cursor.execute call
  insecure_yaml_load	yaml.load → yaml.safe_load
  command_injection	wrap os.system f-string vars in shlex.quote
  subprocess_shell_injection	wrap subprocess(... shell=True) vars in shlex.quote
  eval_injection	eval(...) → ast.literal_eval(...)
  weak_cryptography	hashlib.md5 → hashlib.sha256

• Verification gate — every patch must parse, satisfy an AST assertion proving the specific fix is present, and (when Semgrep is installed) survive a re-scan. Anything short of a verified_fix is not turned into a PR.
• missing_auth_decorator is report-only — a missing auth gate is surfaced for a human, but never auto-injected (auto-injecting an auth decorator cannot know which decorator is correct or whether a route is intentionally public).
• Drift detection — snapshot a repo's security state and detect regressions over time.

Installation

pip install aion-evolve
# or
uv tool install aion-evolve

For local development:

git clone https://github.com/shenxianpeng/aion.git
cd aion
uv sync --group dev --group docs
uv run aion --help

Quick start

Scan and explain (LLM)

scan uses an LLM provider for explanations. Set at least one key:

export OPENAI_API_KEY=your_key      # or ANTHROPIC_API_KEY / DEEPSEEK_API_KEY /
                                    #    QWEN_API_KEY / GEMINI_API_KEY
aion scan ./path/to/repo --output json

Plan and verify a single deterministic repair (no LLM needed)

aion repair ./path/to/file.py \
  --context-file ./context.json \
  --artifact-path ./artifact.json

aion verify --artifact-path ./artifact.json

Auto-update: scan → verify → open PRs

aion auto-update --target ./ --dry-run   # preview, no PRs
aion auto-update --target ./             # live: opens PRs for verified fixes

GitHub Action

AION ships as a reusable GitHub Action:

# .github/workflows/aion.yml
name: AION Auto-Update
on:
  schedule:
    - cron: '0 9 * * 1'  # Weekly on Monday at 09:00 UTC
  workflow_dispatch:

permissions:
  contents: write
  pull-requests: write

jobs:
  auto-update:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v6
        with:
          fetch-depth: 0
      - uses: shenxianpeng/aion@main
        with:
          openai_api_key: ${{ secrets.OPENAI_API_KEY }}
          # or: deepseek_api_key / qwen_api_key / anthropic_api_key / gemini_api_key

Configuration

AION reads a flat .aion.yaml (CLI flags override it):

provider: openai
model: gpt-4.1
ignore_paths:
  - tests/*
  - scripts/generated_*.py
auto_repair_issue_types:
  - raw_sqlite_query
  - hardcoded_secret
  - insecure_yaml_load
  - command_injection
auto_repair_min_confidence: 0.90
open_pull_requests_limit: 5
labels:
  - "aion"
  - "security"
reviewers:
  - "team:security"
assignees:
  - "username"
target_branch: "main"
commit_message_prefix: "[AION]"

Supported LLM providers

Command surface

Scope (what AION does not do)

• It produces patch artifacts and pull requests. It does not hot-patch live production code, and it is not a runtime control plane.
• It is Python-only by design.
• It fixes a deliberately small set of high-confidence issue types. Breadth is a non-goal; trustworthiness is the goal.

Documentation

Published at shenxianpeng.github.io/aion (中文).