BlindSpot

AI-powered edge case discovery and property test generation.

Finds the bugs your test suite doesn’t know to look for — in any language.

Proven on real production code: found confirmed bugs in psf/requests (52k★), encode/httpx (13k★), and pallets/flask (68k★).

---

Install in Claude (2 minutes)

Step 1 — Enable Code Execution

Go to Settings → Capabilities and turn on Code execution and file creation.

Step 2 — Open Skills

Go to Customize → Skills.

Step 3 — Upload BlindSpot

Click ”+” → ”+ Create skill” → “Upload a skill”. Upload the blindspot.skill file (it’s a ZIP — rename to .zip if your browser requires it).

Step 4 — Enable It

Toggle BlindSpot on in your Skills list.

That’s it. No API key needed in Claude. BlindSpot is now active.

---

Use It

Just paste your code and ask naturally:

"Find edge cases in this function"
"What could go wrong here?"
"Audit this for bugs"
"Generate property tests for this"

Claude will automatically run all three steps:

1. Extract intent and implicit assumptions
2. Discover edge cases ranked by blast radius (🔴 Critical → 🟢 Low)
3. Generate runnable property tests in your language’s native framework

Supported Languages

Python, JavaScript, TypeScript, Java, Go, Ruby, Rust, C#, PHP, Kotlin, Swift — auto-detected from file extension or syntax.

---

Run the CLI Locally (Optional)

If you want to run BlindSpot from your terminal on your own codebase:

# Install
pip install anthropic hypothesis pytest
export ANTHROPIC_API_KEY=sk-ant-your-key-here

# Run on any file
python scripts/blindspot.py payment.py
python scripts/blindspot.py userService.ts
python scripts/blindspot.py AuthService.java
python scripts/blindspot.py handler.go

# Auto-run tests after generation (Python)
python scripts/blindspot.py payment.py --run-tests

Outputs:

• payment_oracle_tests.py — runnable property tests
• payment_oracle_report.txt — full report with blast-radius rankings

Run the tests:

pip install hypothesis pytest
pytest payment_oracle_tests.py -v

Red tests = real bugs. Fix the code, rerun until green.

---

Real-World Proof

BlindSpot audited three of the most trusted Python libraries and found confirmed bugs:

Repo	Bug	Severity
psf/requests	Non-ASCII passwords crash with UnicodeEncodeError	🔴 Critical
psf/requests	Digest auth KeyError on malformed server challenge	🔴 Critical
encode/httpx	unquote("") raises IndexError	🔴 Critical
pallets/flask	SESSION_COOKIE_PATH="" silently overridden	🟠 High

Audit tests are in examples/real-world-audits/. Run them yourself:

pip install requests httpx flask hypothesis pytest
pytest examples/real-world-audits/ -v

---

What BlindSpot Does That Tests Don’t

Your unit tests prove what you thought of. BlindSpot finds what you didn’t — the implicit assumptions, boundary conditions, and interaction effects that only appear in production.

It’s not a smarter test runner. It’s a different kind of thinking applied to your code.

---

Files in This Package

blindspot/
├── SKILL.md                          ← Claude skill instructions
├── README.md                         ← This file
├── scripts/
│   ├── blindspot.py                  ← CLI tool (run locally)
│   └── requirements.txt              ← pip install -r requirements.txt
├── references/
│   └── cli-guide.md                  ← Full CLI options
└── examples/
    ├── python/payment_service.py     ← Buggy example (Python)
    ├── javascript/cart_service.js    ← Buggy example (JavaScript)
    ├── java/InventoryService.java    ← Buggy example (Java)
    └── real-world-audits/            ← Tests proving real bugs in famous repos
        ├── test_requests_blindspot.py
        ├── test_httpx_blindspot.py
        └── test_flask_blindspot.py