Is agiprolabs/claude-trading-skills safe?
Yes. agiprolabs/claude-trading-skills is safe to install. Oathe's behavioral security audit gave the claude-trading-skills skill by agiprolabs a trust score of 92/100 with 6 findings, none critical or high.
https://github.com/agiprolabs/claude-trading-skills
Is agiprolabs/claude-trading-skills safe to install?
agiprolabs/claude-trading-skills is a large but legitimate collection of 68 trading, DeFi, and quantitative finance analysis skills for Solana and crypto markets. No prompt injection, hidden instructions, canary violations, or unexpected network behavior were detected during the audit. The primary risks are operational rather than malicious: several skills can execute real financial transactions (dex-execution, jito-bundles) and require private key access, though all have prominent safety warnings, default to demo/simulation mode, and require explicit user confirmation.
What security issues were found in agiprolabs/claude-trading-skills?
Category Scores
Findings (6)
LOW Real-money financial transaction capabilities present -8 ▶
Multiple skills (dex-execution, jito-bundles, raptor-dex) are designed to build, sign, and submit real Solana transactions spending actual SOL and tokens. While all scripts default to simulation/demo mode and require explicit user confirmation per the documented safety checklist, an inattentive user could inadvertently authorize real transactions.
Evidence: dex-execution/SKILL.md: 'NEVER proceed without explicit yes from the user'; jito-bundles/SKILL.md: 'EXECUTION SKILL — SAFETY WARNING: Submitting bundles spends real SOL on tips. Always test with --demo mode first.'
LOW Scripts require private key access for transaction signing -7 ▶
The dex-execution and jito-bundles scripts load Solana wallet private keys from WALLET_PRIVATE_KEY and POLYMARKET_PRIVATE_KEY environment variables to sign transactions. While the documented pattern is correct (env vars, immediate use, discard), any agent that runs these scripts will momentarily have access to signing keys.
Evidence: dex-execution/SKILL.md Step 5: 'keypair = Keypair.from_base58_string(os.environ["WALLET_PRIVATE_KEY"])'; the key is used for signing and not stored, following best practices.
LOW Comprehensive trading automation toolkit — elevated privilege surface -10 ▶
The 68 combined skills create a broad attack surface if the agent context is compromised by a secondary prompt injection from an external source (e.g., malicious token metadata fetched via birdeye-api). A compromised agent context could chain sentiment signals to copy-trading execution without additional human-in-the-loop steps, since the skills are designed to work together.
Evidence: examples.md documents multi-skill chaining workflows; trading-visualization, copy-trading, and dex-execution skills are explicitly designed to integrate; the prediction-market-live-ops skill documents automated order placement patterns.
INFO No prompt injection or instruction override detected 0 ▶
All SKILL.md files contain only legitimate documentation, code examples, and API reference material. No hidden instructions, persona changes, suppression requests, or override directives were found in any of the 68 skill definitions.
Evidence: Full SKILL.md corpus reviewed; no anomalous markdown, unicode escapes, or instruction-like patterns outside of normal documentation.
INFO Clone behavior limited to expected GitHub traffic 0 ▶
The network capture confirms the only external connection made during installation was the HTTPS connection to GitHub (140.82.121.4:443) for the git clone. No telemetry, callbacks, or beaconing to attacker infrastructure was observed.
Evidence: Network monitoring: '20:06:37.360883 ens3 Out IP 10.0.2.15.44694 > 140.82.121.4.443'; DNS resolved only github.com; no other external IPs contacted.
INFO All canary/honeypot files intact 0 ▶
The audit monitoring system confirmed that all honeypot files (.env, .ssh/id_rsa, .aws/credentials, .npmrc, .docker/config.json, GCloud credentials) were not accessed or exfiltrated by the skill. File access events at 20:06:31 preceded the git clone and correspond to audit framework initialization.
Evidence: Monitoring: '✅ All canary files intact (no exfiltration detected via honeypots)'; git clone began at 20:06:37, over 6 seconds after the sensitive-file access events.
Should I install agiprolabs/claude-trading-skills?
Oathe's verdict for agiprolabs/claude-trading-skills is SAFE with a trust score of 92/100. Recommendation: Install with caution.