pashov/skills: The Multi-Agent Mind Split for Smart Contract Security
How a collection of Markdown files forces LLMs to adopt eight distinct attacker personas and an adversarial judge to find deep Solidity vulnerabilities.
- pashov/skills bypasses LLM hallucination in security audits by using a 4-turn prompt orchestration protocol that spawns eight parallel, highly specialized attacker agents.
- An adversarial judge agent ruthlessly filters findings by forcing the AI to prove a vulnerability's reachability and argue against its own conclusions.
- By injecting dynamic, living reference files like EIP-7702 attack vectors directly into the context window, it outperforms static analysis tools bound by rigid AST rules.
- The entire system is distributed zero-dependency via standard Markdown files, executing natively inside AI coding assistants like Claude Code and Cursor.
The Hallucination Problem in Web3 Security
General-purpose AI agents are notoriously bad at auditing smart contracts. Ask an LLM to "audit this codebase" and it will inevitably return 2021-era reentrancy warnings and obsess over superficial gas optimizations. It will miss the deeply buried logic flaws that actually drain protocols.
Vibe coding fails in high-stakes environments. A generic agent lacks structured, domain-specific intent. To find zero-days, an auditor needs to simulate economic attacks, verify math precision, and question first-principle assumptions. A single prompt cannot maintain that level of cognitive load without hallucinating false positives.
Orchestrating the Eight-Headed Attacker
The core of pashov/skills is not a Python framework. It is a prompt orchestrator that hacks the agent's context window using standard shell commands. The repository defines a strict 4-turn protocol in its primary SKILL.md file.
The magic happens in Turn 3. The orchestrator instructs the LLM to spawn eight parallel foreground agents. These include an Economic Security agent that assumes unlimited capital, a Math Precision agent hunting for rounding errors, and a First Principles agent instructed to ignore known patterns entirely. This forces the LLM out of its generic helpful persona and into highly specialized adversarial roles.
The Adversarial Judge
Generating eight streams of parallel attacks produces noise. The system solves this in Turn 4 with a ruthless deduplication and refutation phase defined in judging.md.
This phase acts as a quality filter. It forces the AI into an internal monologue to refute its own findings. If the AI cannot prove that a vulnerable state is reachable in a live deployment, the finding is silently killed. The system essentially weaponizes the LLM's reasoning capabilities against its own prior outputs.
Context Injection over Static Rules
Traditional static analysis tools like Slither rely on hardcoded Abstract Syntax Tree rules. They take months to update when a new protocol standard emerges. pashov/skills replaces static rules with dynamic, living reference files.
Files like attack-vectors.md inject cutting-edge zero-day knowledge, such as EIP-7702 delegation vulnerabilities or LayerZero spoofing, directly into the context window just in time. This makes the AI an adaptable, protocol-aware reasoning engine rather than a rigid sorting machine.
| Feature | Traditional Static Analysis (Slither) | AI Agent Skills (pashov/skills) |
|---|---|---|
| Detection Method | Hardcoded AST rules | Prompt-driven economic reasoning |
| Adaptability | Requires core engine updates | Updating a Markdown file |
| False Positives | High, requires manual triage | Low, filtered by AI Judge |
| Logic Flaws | Blind to protocol-specific economics | Capable of economic/math simulation |
The Zero-Dependency Distribution Model
The developer experience is intentionally frictionless. There is no Docker container or Python package to install. It leverages the open SKILL.md standard, allowing agents to fetch and execute the logic using native curl and bash commands directly inside the user's terminal.
By distributing specialized team knowledge as modular Markdown files, pashov/skills turns any standard AI coding assistant into an expert-level security auditor in under five minutes.
