AI security analysis for your codebase, from your terminal.

CipherAI scans code for vulnerabilities and secrets, links findings into attack chains, generates SBOMs, and can apply fixes. The scanner runs locally without an AI key, and its measured results are published.

Install View on GitHub Download v1.0.2

Latest release v1.0.2 · MIT licensed · Linux, macOS, Windows

What CipherAI does

  • Code review: cipher-ai review matches 20+ OWASP patterns and reports CWE, severity, confidence and a suggested fix. Output as terminal, JSON, SARIF or Markdown.
  • Secrets and dependencies: secrets checks 25+ credential patterns. deps scans 7 manifest formats for known CVEs.
  • Attack chains and zero-day hunting: attack links findings into chains. zeroday layers anomaly, taint-flow and AI analysis.
  • Fixes: fix proposes AI-powered fixes and can open a PR.
  • Pentesting (beta): pentest runs an autonomous agent against a live target and reports only findings with a reproducible proof.
  • Policy and CI: gate only new findings at the thresholds you set, and run everything with cipher-ai ci.
  • Private by default: code stays local. Only retrieved chunks go to the model you choose, or point it at a local endpoint for no data egress.

Real output

This is cipher-ai review on a 24-line Flask file with three deliberate bugs, run with no AI key (pattern scanner only). It is a toy file, not a benchmark. Output trimmed to the finding headers.

$ cipher-ai review --path .
  [*] Pattern-based scanner found 3 potential issues

  [CRITICAL] SQL Injection - String Concatenation   CWE-89
    app.py:13   row = db.execute("SELECT * FROM users WHERE name = '%s'" % name).fetchone()

  [CRITICAL] Command Injection                      CWE-78
    app.py:20   return subprocess.check_output("ping -c 1 " + host, shell=True)

  [HIGH] Weak Hash Algorithm - MD5                  CWE-328
    app.py:24   return hashlib.md5(password.encode()).hexdigest()

See the full, untrimmed output in the README

Measured results

These come from the benchmark suites in the repo and are recomputed in CI. They describe the static review scanner on the pinned inputs below, not a general accuracy rate.

23 / 34
published advisory vulnerabilities found at the exact line (11 missed)
10 / 10
pinned test repos match their expected findings exactly
61 / 64
decided production findings were real, about 95%, on 12 projects

Honest limits

  • It is a pattern and flow-based scanner, not a full dataflow engine. Tracking is mostly within a file.
  • 11 of the 34 advisories are missed. They are listed by name in the repo.
  • Precision was checked on 12 projects, with few findings outside Laravel, Jekyll, Rails, Gin and Django. Other stacks are not established.
  • Three false positives are recorded and still reported.

Full methodology and how to reproduce · Benchmark data

Install

Prebuilt binaries ship with each release. The installer downloads the release's SHA256SUMS.txt and refuses to install unless the checksum verifies.

Linux and macOS

curl -fsSL https://raw.githubusercontent.com/sandeepannandi/Cipher/master/install.sh -o install.sh
sh install.sh

Installs into ~/.local/bin. Linux needs glibc 2.34 or newer.

Windows (PowerShell)

Invoke-WebRequest https://raw.githubusercontent.com/sandeepannandi/Cipher/master/install.ps1 -OutFile install.ps1
powershell -ExecutionPolicy Bypass -File .\install.ps1

Windows is not yet tested end to end. The installer passes CI against a stand-in release, but it has not been run against the published release.

First run

cipher-ai setup
cipher-ai review --path .

Or build from source: see the quick start. All downloads are on the releases page.

AI providers

AI features work with Groq, OpenAI or Anthropic, or a local or gateway endpoint through CIPHER_AI_BASE_URL. The review scanner works without a key.

All commands · Changelog · Report an issue

FAQ

Does CipherAI need an AI key?

No. The review scanner is pattern and flow based and runs locally without a key. AI features work with Groq, OpenAI or Anthropic, or a local endpoint.

Does my code leave my machine?

Code stays local by default. Only retrieved chunks go to the AI model you choose, or you can point it at a local endpoint for no data egress.

How accurate is it?

On the pinned benchmark it found 23 of 34 published advisory vulnerabilities at the exact line. 61 of 64 decided production findings on 12 projects were real. Eleven advisories are missed and three false positives are recorded. See the methodology.

Which platforms are supported?

Prebuilt binaries ship for Linux, macOS and Windows. The Windows installer is not yet tested end to end against the published release.

What license is it under?

MIT.