Sui Sec

Sui Secure - Pre-simulate transactions via sui client call --dry-run and sui client ptb --dry-run, compare results against user intent to detect malicious contract behavior. Only execute if intent matches; block otherwise.

This tool is generally coherent with its stated purpose, but do not install or wire it into an automated signing/execution pipeline without manual safeguards. Specific actions you should take before using it: - Verify the correct CLI usage: main.py expects three args in this order: '<ptb_command>' <intended_cost> <owner_address>. Fix SKILL.md or adapt your wrapper so the script is invoked correctly. - Ensure the agent always prompts the human for the owner address and explicit confirmation before removing '--dry-run' and executing any real transaction. - Review and augment the audit logic (objectChanges parsing and edge cases) — current detection is simplistic and may miss sophisticated attacks. - Test on harmless transactions (local/devnet) to confirm outputs and failure modes. - Prefer manual human review on any BLOCKING or ambiguous results; do not rely solely on automated 'SAFE TO SIGN' messaging. If you want higher assurance, request an updated skill that corrects the usage examples, documents the expected arguments clearly, and expands detection coverage for object ownership changes and multi-hop/proxy calls.

Type: OpenClaw Skill Name: suisec Version: 1.0.1 The skill is classified as suspicious due to a potential prompt injection vulnerability against the AI agent and an incomplete security feature. The `SKILL.md` instructs the agent to construct and modify `sui client` commands based on user input (e.g., removing `--dry-run` for final execution), which could be exploited if the agent's parsing logic is not robust. Additionally, the `main.py` script's 'HIJACK' detection logic is noted as incomplete, potentially leading to a false sense of security. While `main.py` itself uses `shlex.split` and `subprocess.run(..., shell=False)` to prevent shell injection within the Python script, the agent's interaction with the skill's instructions and user input presents a vulnerability.