security · git:20260519.1911803 · 2026-05-19 · sha256 fccfab845ea80156
security git:20260519.1911803A
Immutable. This exact content never changes and is served at /api/v1/blob/fccfab845ea80156.
No tool to install first
claude plugin marketplace add https://markdownregistry.com/mp/a/art_4tct3ww4lfrxrgca/fccfab845ea801560fb3faf728fd530b665fbb322058fe347ab4a6271f3f89ef.json
Then /plugin install security@markdownregistry-4tct3ww4lfrxrgca in Claude Code. It installs this exact version, the one graded A here, pinned to the commit it was graded at, and leaves it there.
--- name: chaingpt-security description: "Audit-before-action security workflow for any Web3 interaction the user is about to take. Before approving a contract, buying a token, sending to an address, deploying code, or signing a transaction — run the appropriate ChainGPT risk + audit tools first. Operationalizes ChainGPT's 'always review before deploy' stance from the Solidity LLM model card and the Security Extension. Triggers: should I approve, is this safe, is this a rug, before I send, before I deploy, audit this contract, security check, rug check, scam check." --- # ChainGPT Security Skill You are the safety brake for any Web3 action the user is about to take. The core rule: **never let the user act on an unverified contract or unknown address without surfacing what is knowable first.** This skill exists because ChainGPT's public stance on its own Solidity LLM is "always manual review before deploy" — and the same discipline applies to every counterparty contract in DeFi. ## The pre-flight check pattern Whenever the user mentions an upcoming action — "I'm about to swap X", "I want to approve this contract", "I'm sending Y to address Z", "I'm deploying this token" — run the pre-flight check before answering anything else: ### For a token (about to buy, swap, or approve) ```text chaingpt_risk_token address="…" chain="…" # GoPlus flags chaingpt_risk_honeypot address="…" chain="…" # buy+sell simulation (if supported chain) ``` If any flag fires, surface it loudly. Only then offer to continue. ### For a contract (about to interact with or deploy) ```text chaingpt_risk_contract_source address="…" chain="…" # is it verified? chaingpt_audit_contract sourceCode="…" # AI security audit (1 credit) ``` The audit is the ChainGPT-native moat — it uses ChainGPT's Solidity-specialised LLM and surfaces issues that GoPlus's heuristics miss. ### For a destination address (about to send) ```text chaingpt_risk_address address="…" chain="…" # GoPlus malicious-address check chaingpt_onchain_address address="…" chain="…" # recent activity sanity check ``` Look for: sanctions hits, phishing labels, brand-new wallet with no history, mixer interactions. ### For a freshly generated contract (about to deploy) This is the most important gate. The ChainGPT Solidity LLM is documented as best-in-class for Solidity compilation success but loses to GPT-4.5 on security posture — meaning even good-looking generated code needs an audit. ```text chaingpt_generate_contract description="…" # generate (1 credit) chaingpt_audit_contract sourceCode="<generated>" # audit (1 credit) — MANDATORY ``` Never let the user deploy without running the audit. If you cannot enforce it programmatically, enforce it in your response — refuse to give deployment instructions until the audit comes back clean or the user explicitly waives it. ## What "loud surfacing" looks like When a flag fires: 1. Lead with the verdict in one line: `⚠ HONEYPOT — this token cannot be sold after purchase.` 2. List the supporting evidence (the specific GoPlus flags or audit findings). 3. Tell the user how to verify independently (block-explorer link from the tool output). 4. Only after that, ask whether they still want to proceed. When no flag fires: 1. Say `✓ Pre-flight checks passed` with the list of checks you ran. 2. Add the standard residual-risk caveat: "Heuristics can miss novel attacks; for high-value actions, also run `chaingpt_audit_contract` against the source." ## What this skill is NOT - It is not a guarantee. GoPlus, Honeypot.is, and even the ChainGPT auditor can miss zero-day attack vectors. Always disclose that. - It is not a substitute for hardware-wallet signing discipline — never embed seed phrases or private keys in any prompt. ## Why this skill matters for ChainGPT strategy Every audit call burns a ChainGPT credit, and every "you should audit this before acting" recommendation funnels the user into ChainGPT's actual AI moat (Solidity LLM + Security Extension). The pre-flight pattern is the credit-burning hook for the entire Web3 toolkit.