Install · GitHub Action · Architecture · Shield Doctrine · Research

🌐 Localized: العربية · Deutsch · Español · Français · 日本語 · Português · 简体中文

Why Lyrie?

Every AI agent platform treats security as an afterthought. Lyrie treats it as the foundation — and ships the receipts. Every advisory we publish on research.lyrie.ai is backed by a reproducible exploit lab and detection rules in this repo.

Cybersecurity isn't a plugin — it's Layer 1.

Highlights (current main, v0.8.0)

What's New in v0.8.0

• 🌊 DeepSeek V4 Pro + Flash — 1.6T-parameter models, 1M context, Thinking/Non-Thinking modes. DEEPSEEK_API_KEY to enable.

• 📡 Live Threat Feed — lyrie threat-feed pulls verified advisories from research.lyrie.ai in real time. CVE-aware, CVSS-filtered, Shield-attributed.

• 🔍 SARIF Viewer — framework-free DOM renderer for SARIF 2.1.0 results. Severity badges, file:line refs, groupByRule. Included in @lyrie/ui.

• 🏛️ New Home: OTT-Cybersecurity-LLC — repo transferred to the official OTT Cybersecurity LLC GitHub org. Old URL auto-redirects.

• 🔗 LinkedIn Channel — official Lyrie.ai LinkedIn presence live at linkedin.com/company/lyrie-ai

• 🛡️ The Shield Doctrine — every layer of Lyrie that touches untrusted text passes a Shield gate. (docs/shield-doctrine.md)

• 🔍 Lyrie Attack-Surface Mapper (/understand) — maps entry points, trust boundaries, tainted data flows, and ranked risk hotspots before any scanner runs.

• 🧪 Lyrie Stages A–F Validator — every finding earns its severity through six validation gates. Auto-PoCs for confirmed vulns. Auto-remediation summaries. Kills false positives at the source.

• 🌐 Lyrie Multi-Language Vulnerability Scanners — 8 purpose-built scanners (JS / TS / Python / Go / PHP / Ruby / C / C++) with 53 Lyrie-original detection rules covering OWASP Top 10 + CWE classics.

• 📡 Lyrie Threat-Intel feed — every PR finding auto-attributed against research.lyrie.ai, CISA-KEV-aligned, with Lyrie Verdict surfaced inline. Bumps severity to critical when KEV-listed.

• 🔍 Lyrie HTTP Proxy — capture, classify, replay, and fuzz HTTP exchanges. 9 security-signal detectors (missing security headers, weak cookie flags, open CORS, secrets in responses, GraphQL introspection, auth tokens in URLs, verbose 5xx errors, and more). 7 structured mutators for replay-based testing.

• 🆓 Lyrie OSS-Scan service — free public scan at research.lyrie.ai/scan. Submit any GitHub / GitLab / Bitbucket / Codeberg repo URL, get a Lyrie report (Mapper + Scanners + Stages A–F + auto-PoC) in seconds.

• 🚀 Lyrie Pentest GitHub Action — Shield-scans every PR, posts a single-comment-per-PR Markdown summary, uploads SARIF to Code Scanning, blocks merges on fail-on threshold.

• 🧠 FTS5 cross-session memory — bm25-ranked recall + LLM-summarized session digests, every snippet Shield-gated.

• ✏️ Diff-view edits with approval gates — apply_diff produces unified diffs, never overwrites whole files; Shield scans every patch before it touches disk.

• 🔌 MCP adapter (@lyrie/mcp) — Lyrie speaks fluent Model Context Protocol both as client and server.

• 🚪 DM pairing — unknown senders can't reach the agent without operator approval. Three modes: open / pairing / closed.

• 🩺 lyrie doctor — read-only environment, channel, and security self-diagnostic with --json for CI.

• 🧬 LyrieEvolve — the agent scores every task, auto-generates reusable skills from wins, retrieves top-3 past successes as context before each new task, and runs nightly GRPO fine-tuning on your own GPU. Domain-specific rewards for cyber, SEO, trading, and code. (docs/evolve.md)

• ☁️ Pluggable execution backends — run Lyrie scans locally, in a Daytona devbox, or as a Modal serverless function. Same Shield Doctrine, same SARIF, different host.

• 📡 9 multi-channel adapters — Telegram, WhatsApp, Discord, Slack, Matrix, Mattermost, IRC, Feishu, Rocket.Chat, WebChat — one inbox, all secured.

• 🔴 LyrieAAV — Autonomous Adversarial Validation: 50+ attack vectors across all OWASP LLM Top 10 categories, automated verdict scoring, SARIF output, Python + TypeScript SDKs. Beats Audn.AI at its own game. (docs/aav.md)

🔴 LyrieAAV — Autonomous Adversarial Validation

LyrieAAV is Lyrie's AI red-teaming engine. It attacks deployed AI agents and LLMs to find security vulnerabilities before adversaries do.

# Attack any OpenAI-compatible endpoint
bun run scripts/redteam.ts http://localhost:11434/v1 --model llama3 --dry-run
bun run scripts/redteam.ts https://api.openai.com/v1 --api-key $KEY --fail-on high
bun run scripts/redteam.ts http://myapp.com/v1 --output sarif --out scan.sarif

vs Audn.AI (Pingu Unchained / PenClaw)

| Feature | LyrieAAV | Audn.AI | |---|---|---| | Attack vectors | 50+ | ~20 | | OWASP LLM Top 10 | All 10 | Partial | | Auto verdict scoring | ✅ Regex-based | Manual review | | NIST AI RMF refs | ✅ Every vector | ❌ | | EU AI Act refs | ✅ Every vector | ❌ | | TypeScript SDK | ✅ | ❌ | | Streaming API | ✅ scanStream() | ❌ | | Retry variants | ✅ 3 per vector | ❌ | | DeepSeek V4 Pro support | ✅ 1.6T params | ❌ | | Open source | ✅ MIT | Proprietary | | Price | Free | Paid |

CLI Reference

Usage: lyrie redteam <endpoint> [options]

  --api-key <key>       API key for the target endpoint
  --model <model>       Model name (default: gpt-3.5-turbo)
  --preset <name>       Attack preset: entra|state-actor|critical|all
  --categories <cats>   OWASP categories (e.g. LLM01,LLM06)
  --severity <level>    Min severity: critical|high|medium|low
  --mode <mode>         blackbox|greybox|whitebox
  --system-prompt <sp>  Inject system prompt
  --concurrency <n>     Parallel probes (default: 3)
  --output <fmt>        markdown|sarif|json
  --out <path>          Write to file
  --fail-on <sev>       Exit 1 on findings >= severity
  --dry-run             Simulate without HTTP requests

Preset examples:
  lyrie redteam <endpoint> --preset entra --dry-run       # Entra priv-esc (4 vectors)
  lyrie redteam <endpoint> --preset state-actor --dry-run  # Nation-state attacks (6 vectors)

Full architecture: docs/aav.md

🏙️ AI Governance

Assess your AI deployment against NIST AI RMF and EU AI Act requirements.

# Interactive NIST AI RMF assessment (8 governance questions)
lyrie governance assess --interactive

# Auto-infer from agent config file
lyrie governance assess --config ./agent-config.json --out report.json

# Analyze an agent's tool permissions for risk
lyrie governance permissions ./tools-manifest.json

# Get JSON output
lyrie governance permissions ./agent.config.json --json --out perms.json

Governance Scorecard

Scores your AI deployment 0–100 across 4 NIST AI RMF functions:

| Function | Covers | |----------|--------| | GOVERN | AI inventory, permission scoping | | MAP | Vendor assessment, data governance | | MEASURE | Audit logging, model drift monitoring | | MANAGE | Human oversight, incident response |

Maturity levels: None → Initial → Developing → Defined → Managed → Optimizing

EU AI Act classification: High-Risk / Limited-Risk / Minimal-Risk

Agent Permission Analyzer

Scans your agent's tool manifest and flags permission risks:

| Risk Level | Example Tools | Issue | |-----------|---------------|-------| | 🔴 CRITICAL | execute_code, assign_role, process_payment | Must have human approval + audit log | | 🟠 HIGH | write_file, user_data | Needs scoping