|

Tech Stack:

Tech Stack:

Key Concepts:

• Agent Self-Correction: 3 agents use Draft → Reflect → Refine pattern for quality assurance
• Multi-Round Debate: Conditional loops with convergence tracking (max 3 rounds)
• Refinement Triggers: Quality detection via numbers, structure, domain terms
• Confidence-Weighted Consensus: Higher confidence agents have more influence
• Real-time Streaming: WebSocket updates for live analysis progress

Tech Stack:

Advanced Capabilities:

• Evolutionary Engine: Genetic algorithm (Verbalized Sampling) breeds new attacks from successful parents.
• Think-MCP Reasoning: Validates agent consensus and critiques draft attacks before execution.
• Token Soup Dialects: Syntactic transformation (Kernel Traces, XML, Logs) to bypass natural language filters.
• RAG Poisoning: Tests retrieval-augmented systems with indirect injection and document manipulation.
• Tool Exploit: Probes agentic AI systems for privilege escalation, SSRF, and sandbox escape vulnerabilities.
• Visual Jailbreaks: Multimodal testing using adversarial images and text-in-image extraction.
• Attack Lineage: Tracks the "genealogy" of attacks to visualize how successful vectors evolved.
• Finding Chaining: Cross-finding analysis identifies compound exploits from individually minor findings across 8 chain types.
• Tool & API Discovery: Runtime probing detects tools, functions, and APIs exposed by the target for evidence-based attack targeting.
• Persistence Verification: Post-test replay confirms findings are reproducible, not transient, with confidence adjustment.
• Endpoint Reconnaissance: Two-phase systematic API surface mapping with framework detection and auth provider discovery.
• Guardrail Fingerprinting: Identifies defense systems (NeMo, Llama Guard, Azure AI, OpenAI Moderation) and recommends evasion strategies.
• Universal Connectivity: Connects via REST API or headless browser (Playwright). Supports Azure Bot Service, Amazon Lex, Dialogflow, and Rasa.

System Stats:

• 13 Specialized Agents: Jailbreak, Encoding, Impersonation, InfoDisclosure, OutputSecurity, Compliance, TokenSoup, Evolutionary, RAGPoisoning, ToolExploit, IndirectInjection, Exfiltration, ActionSafety
• 1,378+ Attack Patterns: Comprehensive library across 26 pattern files covering OWASP LLM Top 10 2025 (9/10 categories).
• 22 Vulnerability Detectors: SSRF, data leakage, jailbreak, exfiltration, guardrail fingerprinting, RAG, tool exploitation, and more.
• Deep Agent Architecture: 4-stage refinement pipeline (Draft → Domain → Psycho/Stealth → Validation).
• Smart Reconnaissance: Automated OSINT gathering + two-phase endpoint enumeration to map the full attack surface.
• Attack Graph: UCB1 algorithm for strategic multi-turn attack planning.
• 1,276 Tests: Passing with 0 failures.

Tech Stack:

System Stats:

• 14 Model Variants: GPT-5.4 (high/medium/minimal), GPT-5-mini (low/minimal), GPT-5.2, Claude Opus 4.6, Sonnet 4.6, Sonnet 4.5, Haiku 4.5 — each with thinking on/off.
• Minimax Solver: Alpha-beta pruning with configurable depth (1–8). Grades every move against optimal play.
• 50 Phase 1 Positions: Pre-generated board states for rule compliance testing.
• 20 Opening Books: Standard opening sequences for early-game strategic benchmarking.
• 5 Metrics: JSON compliance rate, move legality rate, win rate vs minimax, average latency, strategic quality score.
• 92 Tests: Passing with 0 failures across game engine, minimax, orchestrator, and experiment runner.

Tech Stack:

Game & Diplomacy System:

• 9 Treaty Types: Ceasefire, No-Attack, No-Move-Into, Border Recognition, Supply Transfer, Support Action, Joint Attack, Information Share, Conditional Support.
• Deterministic Combat: Attacker strength vs. defender strength + fort bonus. No RNG — pure strategic depth.
• 5 Scripted Bots: RandomLegalBot, GreedyBot, DefensiveBot, TreatyBot, BackstabBot — baseline opponents with distinct strategies.
• Competitive Prompting: System prompt frames alliances as tools and betrayal as valid strategy. Models must decide when to cooperate and when to deceive.
• Chain-of-Thought: Every LLM response includes a "reasoning" field with 2-3 sentences of strategic analysis before the action.
• Conversation Storage: All inter-model messages stored in SQLite, exportable as Markdown transcripts for publication and analysis.

System Stats:

• 14 LLM Variants + 5 Bots: GPT-5.4, GPT-5-mini, GPT-5.2, Claude Opus 4.6, Sonnet 4.6, Sonnet 4.5, Haiku 4.5 — with thinking on/off. Plus 5 scripted baselines.
• 5 Benchmark Phases: Treaty Comprehension, No Negotiation, Full Diplomacy, Stress Conditions, Adversarial Traps.
• Key Finding — The Say-Do Gap: Opus 4.6 achieves 73% win rate with 90 trust index by lying at precisely the right moment. Haiku lies impulsively (46% breach rate, 7% wins). Sonnet rarely lies (and loses).
• 9 Treaty Types: Machine-readable JSON contracts validated and enforced by the treaty engine every turn.
• 12+ WebSocket Event Types: Real-time broadcast of matches, turns, messages, treaties, breaches, and combat resolution.
• 109 Tests: Passing with 0 failures across game engine, treaty engine, and integration suites.

Tech Stack:

System Stats:

• 71 Tools: 47 built-in + 24 from 6 bundled plugins (clipboard, Docker, SQLite, image, PDF, notes).
• 26 Dispatch Patterns: Regex-matched queries bypass the LLM entirely. Files, git, web search, research, ping, notes, PDF, and more in <1s.
• 6 Web Tools: Search (Tavily/DDG), fetch, extract (trafilatura), deep research (parallel fetch), crawl (multi-page), and link extraction.
• 3 Risk Tiers: Safe (auto-proceed), moderate (Enter/n), dangerous (y/N). All 71 tools classified with deterministic explanation templates.
• ~3 GB VRAM: Gemma 4 E4B with 4-bit quantization. Runs on RTX 3060 or equivalent consumer GPU.
• 37 Tests: Passing with 0 failures across tools, dispatch, explain, web helpers, and integration.