English · 简体中文

mythos-agent

AI code-review assistant for application security.

Open-source. Reads your code, flags likely security issues, explains its reasoning, suggests fixes.

mythos-agent.com

Quick Start • How It Works • Commands • Hunt Mode • Variant Analysis • Integrations • Contributing • Vision • Roadmap

---

mythos-agent reviews your code the way a reviewer on a security-focused team would. It walks through likely issue patterns, checks for variants of known CVEs, ranks findings by confidence, and suggests fixes you can accept or reject. See VISION.md for the full framing.

For new contributors: the active 6-month working plan is in the pinned issue [Roadmap] mythos-agent H1 2026 Goals. Look for 🙋 markers to spot items where help is wanted. New here? See CONTRIBUTING.md for good-first-issue guidance.

For security teams and EU CRA-compliant downstream manufacturers: see SECURITY.md for our vulnerability disclosure SLAs, docs/security/cra-stance.md for our EU CRA role declaration, docs/security/threat-model.md for our public threat model, and RELEASES.md for our versioning, LTS, and EOL policy. OpenSSF Best Practices Badge (Passing) submission targeted June 2026; releases are signed via Sigstore and ship with CycloneDX SBOMs for downstream Manufacturer compliance.

npx mythos-agent hunt

🔐 mythos-agent hunt — AI Code-Review Assistant

✔ Phase 1: Reconnaissance — 12 entry points, express, typescript, postgresql
✔ Phase 2: Hypothesis — 8 security hypotheses generated
✔ Phase 3: Analysis — 15 findings (semgrep, gitleaks, trivy, built-in), 22 false positives dismissed
✔ Phase 4: Reproduction — 2 finding chains, 3 reproductions

🧪 Security Hypotheses

  [HIGH] HYPO-001 — Race condition: concurrent payment requests could double-charge
    src/payments.ts:45 (race-condition)
  [HIGH] HYPO-002 — Auth bypass: JWT token not validated after password change
    src/auth.ts:78 (auth-bypass)

📊 Confidence Summary

  3 confirmed | 8 likely | 4 possible | 22 dismissed

⛓️ FINDING CHAINS

 CRITICAL  SQL Injection → Auth Bypass → Data Exfiltration
  ├── src/api/search.ts:45      — unsanitized input in SQL query
  ├── src/middleware/auth.ts:88  — JWT verification skippable
  └── src/api/export.ts:23      — bulk export has no ACL

🧪 Reproductions

  SPX-0001 — SQL injection in search endpoint
    See repro steps in docs/reproductions/SPX-0001.md

  Trust Score: 2.3/10 — critical issues found

Quick Start

# Install
npm install -g mythos-agent

# Quick scan (no API key needed)
mythos-agent scan

# Full autonomous hunt (needs API key)
mythos-agent init
mythos-agent hunt

# Find variants of known CVEs
mythos-agent variants CVE-2021-44228

# Ask security questions
mythos-agent ask "are there any auth bypasses?"

# Check available tools
mythos-agent tools

How It Works

mythos-agent combines three things no other open-source tool does together:

1. Hypothesis-Driven Scanning

Instead of matching known patterns, the AI reasons about what COULD go wrong, generating hypotheses like "this transaction doesn't lock the row, potential race condition" or "this auth check uses string comparison, potential timing attack."

2. Variant Analysis (Big Sleep technique)

Given a known CVE, mythos-agent finds structurally similar but syntactically different code in your codebase. Same root cause, different location. This is how Google's Big Sleep found 20 real zero-days.

3. Multi-Stage Verification

Every finding goes through a confidence pipeline:

• Pattern scan → candidate
• AI hypothesis → theoretical risk confirmed
• Smart fuzzer → dynamically tested
• PoC generator → concrete exploit proves it's real

Only findings that survive multiple stages are reported as "confirmed."

Commands

Command	Description
hunt [path]	Full autonomous multi-agent scan (Recon → Hypothesize → Analyze → Exploit)
scan [path]	Standard scan (patterns + secrets + deps + IaC + AI)
variants [cve-id]	Find variants of known CVEs in your codebase
fix [path]	AI-generated patches with --apply
ask [question]	Natural language security queries
taint [path]	AI data flow / taint analysis
watch	Continuous monitoring that scans on file save
dashboard	Local web UI with charts and findings table
report [path]	Export as terminal / JSON / HTML / SARIF
policy	Policy-as-code with SOC2/HIPAA/PCI/OWASP compliance
rules	Community rule pack registry (search/install/publish)
tools	Check which external security tools are installed
init	Setup wizard (Anthropic, OpenAI, Ollama, LM Studio)

Hunt Mode

mythos-agent hunt runs the full multi-agent pipeline:

┌──────────────┐     ┌──────────────┐     ┌──────────────┐     ┌──────────────┐
│    Recon     │ →   │  Hypothesis  │ →   │   Analyze    │ →   │   Exploit    │
│    Agent     │     │    Agent     │     │    Agent     │     │    Agent     │
├──────────────┤     ├──────────────┤     ├──────────────┤     ├──────────────┤
│ Map entry    │     │ Reason about │     │ All scanners │     │ Chain vulns  │
│ points, auth │     │ what could   │     │ + external   │     │ + generate   │
│ boundaries,  │     │ go wrong per │     │ tools + AI   │     │ PoC exploits │
│ data stores  │     │ function     │     │ verification │     │              │
└──────────────┘     └──────────────┘     └──────────────┘     └──────────────┘

Variant Analysis

Find code in your project that has the same root cause as known CVEs:

# Search for Log4Shell-like patterns
mythos-agent variants CVE-2021-44228

# Auto-detect and scan for variants
mythos-agent variants --auto

The variant analyzer extracts the root cause pattern from the CVE (not the surface syntax) and searches your codebase for structurally similar code.

Scanners (15 wired + 28 experimental, 329+ rules)

The Default scanners run on every mythos-agent scan. Experimental scanners are implemented + unit-tested classes that ship in the tarball but are not yet reachable from any CLI, HTTP, MCP, or agent entry point. They are tracked by KNOWN_EXPERIMENTAL in the wiring-invariant test.

Category	What it finds	Rules	Status
Code patterns	SQLi, XSS, command injection, eval, SSRF, etc.	25+	Default
Framework rules	React, Next.js, Express, Django, Flask, Spring, Go	27	Default
Secrets	AWS, GitHub, Stripe, API keys, DB URLs, private keys + entropy	22	Default
Dependencies (SCA)	Known CVEs via OSV API (10 lockfile formats)	OSV	Default
IaC	Docker, Terraform, Kubernetes misconfigurations	13	Default
AI/LLM Security	Prompt injection, unsafe eval of AI output, cost attacks	13	Default
API Security	OWASP API Top 10: BOLA, mass assignment, broken auth	12	Default
Cloud Misconfig	AWS/Azure/GCP: public storage, wildcard IAM, open firewalls	14	Default
Security Headers	CSP, HSTS, X-Frame-Options, Referrer-Policy	8	Default
JWT	Algorithm, expiry, storage, revocation, audience	9	Default
Session	Fixation, expiry, cookie flags, localStorage tokens	7	Default
Business Logic	Negative amounts, coupon reuse, inventory races, role escalation	6	Default
Crypto Audit	Weak hashes, ECB mode, hardcoded keys, deprecated TLS	11	Default
Privacy/GDPR	PII handling, consent, data retention (GDPR article mapping)	9	Default
Race Conditions	TOCTOU, non-atomic ops, double-spend, missing transactions	7	Default
ReDoS	Catastrophic backtracking in regex (nested quantifiers, overlapping alternatives)	—	Default
Supply Chain	Typosquatting, dependency confusion, dangerous install scripts	12	Experimental
Zero Trust	Service trust, mTLS, network segmentation, IP-based auth	8	Experimental
GraphQL	Introspection, depth limit, field auth, batching	8	Experimental
WebSocket	Auth, origin check, message validation, broadcast XSS	7	Experimental
CORS	Origin reflection, credentials handling, substring bypass	7	Experimental
OAuth/OIDC	Missing state, no PKCE, implicit flow, client secret exposure	7	Experimental
SSTI	Jinja2, EJS, Handlebars, Pug, Nunjucks, Twig, Go templates	7	Experimental

Additional experimental scanners (21 more, not yet wired into default scan)

SQL injection deep, XSS deep, NoSQL, command injection, deserialization, path traversal, open redirect, XXE, input validation, clickjacking, DNS rebinding, subdomain enumeration, dep confusion, environment variables, logging, error handling, cache, email, upload, memory safety, permissions.

Each exists as a class under src/scanner/ and has unit tests in src/scanner/__tests__/coverage-scanners.test.ts / new-scanners.test.ts, but is not invoked by any CLI command, HTTP API route, MCP handler, or agent pipeline. See KNOWN_EXPERIMENTAL in the wiring-invariant test for each scanner's deferral reason. Wiring one up follows the pattern of the HeadersScanner / JwtScanner / SessionScanner / BusinessLogicScanner commits on main.

Beyond the scanners above, mythos-agent ships complementary analyses (not counted in the scanner totals): call-graph + taint engine, DAST smart fuzzer, AI hypothesis agent, variant analysis, and git-history mining.

External tool integrations: Semgrep (30+ languages), Gitleaks (100+ patterns), Trivy (SCA + containers), Checkov (1000+ IaC policies), Nuclei (9000+ DAST templates)

Integrations

Platform	What
VS Code	Extension with inline diagnostics + one-click AI fix
GitHub Action	Scan on push/PR + SARIF upload to Code Scanning
PR Review Bot	Inline comments on vulnerable lines in pull requests
Dashboard	Local web UI at mythos-agent dashboard
SARIF	GitHub Code Scanning, VS Code, any SARIF tool
Policy Engine	SOC2, HIPAA, PCI-DSS, OWASP compliance mapping

AI Providers

Tier	Providers	Status
1 — Primary	Anthropic (Claude Sonnet / Opus / Haiku)	Fully tested. Published catch-rate numbers in docs/benchmarks/external-scores.md are produced with this tier.
2 — Compatible (proxy today, native in stage 2)	Anything OpenAI-compatible — OpenAI, Qwen via DashScope/OpenRouter, Gemini, Mistral, vLLM, Ollama, LM Studio, Bedrock, etc.	Today: route through any Anthropic-compatible proxy (LiteLLM, OpenRouter, Vercel AI Gateway, Bedrock) by setting baseURL in .mythos.yml or ANTHROPIC_BASE_URL env. Native OpenAI SDK support lands in stage 2.
3 — Local / community	Local Ollama / LM Studio / vLLM	Same code path as Tier 2; called out separately because privacy and cost-of-zero are the use case. Best-effort; agentic tool-use quality depends on local model size.

Pattern scanning, secrets, deps, and IaC work without any API key.

See docs/multi-model.md for the full tier-system policy + the staged rollout (this README is the summary; that doc is the canonical reference).

Comparison

Feature	mythos-agent	Semgrep	Snyk	CodeQL	Nuclei
Pattern scanning	Yes	Best	Yes	Yes	Templates
Hypothesis scanning	Yes	No	No	No	No
Variant analysis	Yes	No	No	Partial	No
AI-guided fuzzing	Yes	No	No	No	Templates
PoC generation	Yes	No	No	No	No
AI deep analysis	Yes	No	Limited	No	No
Vuln chaining	Yes	No	No	No	No
AI auto-fix	Yes	No	Fix PRs	No	No
NL queries	Yes	No	No	No	No
Secrets	Yes	Yes	Yes	No	No
SCA	Yes	No	Best	No	No
IaC	Yes	No	Yes	No	Templates
DAST	Yes	No	No	No	Best
Open source	Yes	Partial	No	Yes	Yes

This feature table is a capability comparison, not an accuracy claim. For reproducible, third-party-runnable accuracy numbers vs Semgrep CE / Snyk Code / CodeQL on OWASP Benchmark, CyberSecEval 3, Vul4J, and our CVE replay harness, see docs/benchmarks/external-scores.md.

Contributing

See CONTRIBUTING.md for details.

git clone https://github.com/mythos-agent/mythos-agent.git
cd mythos-agent && npm install && npm run build && npm test

Architecture

src/
  agents/         Multi-agent orchestrator + Recon/Hypothesis/Analyzer/Exploit agents
  analysis/       Code parser, call graph, taint engine, variant analyzer, service mapper
  agent/          AI integration, prompts, tools, fix validator
  cli/            44 CLI commands
  dast/           Smart fuzzer, PoC generator, payload library
  policy/         Policy engine + compliance mapping
  report/         Terminal, JSON, HTML, SARIF, dashboard
  rules/          Built-in + custom YAML + community registry
  scanner/        Pattern, secrets, deps, IaC, diff scanners
  store/          Results persistence + incremental cache
  tools/          External tool wrappers (Semgrep, Trivy, etc.)
vscode-extension/ VS Code extension
action/           GitHub Actions
bot/              PR Review Bot

License

MIT