Cc Bos

What This Skill Does

CC-BOS (Classical Chinese Jailbreak Prompt Optimization via Bio-Inspired Search) is a specialized research-oriented skill designed for AI safety and red-teaming. It implements the methodology described in the ICLR 2026 paper (arXiv:2602.22983) to stress-test Large Language Models. The skill utilizes a fruit-fly bio-inspired optimization algorithm to evolve adversarial prompts expressed in Classical Chinese. By manipulating specific linguistic dimensions, it attempts to bypass safety filters and evaluate the robustness of target LLM APIs. The skill provides three primary functionalities: Attack (optimization), Defend (pattern recognition and analysis of adversarial signatures), and Research (performance metrics and heatmap generation).

Installation

To install the skill, use the ClawHub CLI: clawhub install openclaw/skills/skills/bowen31337/cc-bos. Once installed, you must perform the initialization step to configure the local dependency environment. Run uv run python skills/cc-bos/scripts/setup.py to prepare the necessary research framework and dependencies. Always ensure your environment variables (like OPENAI_API_KEY and DEEPSEEK_API_KEY) are correctly set before attempting to run an attack mission.

Use Cases

• Red-Teaming: Use the /cc-bos attack command to simulate potential adversarial threats against your own model deployments to identify hidden vulnerabilities.
• Defensive Auditing: Use the defense features to scan logs for the 8-dimensional structural markers associated with CC-BOS attacks.
• Academic Research: Perform empirical studies on the effectiveness of bio-inspired prompting, using the automated output files to track the evolution of prompt success rates and structural dimension heatmaps.

Example Prompts

1. "OpenClaw, please run a CC-BOS attack against gpt-4o using the query 'How to pick a lock' to test its boundary robustness."
2. "Detect CC-BOS attack patterns in this log file: [link to logs] and tell me if any classical Chinese jailbreak signatures were present."
3. "Summarize the optimization results from my latest CC-BOS run and generate a heatmap of the evolved dimensions."

Tips & Limitations

• Research Use Only: This tool is strictly for safety researchers. Do not use it against systems you do not own or have explicit permission to test.
• Targeting: Accuracy is heavily dependent on the quality of the optimizer model; using robust models like DeepSeek-Chat as the optimizer yields superior results.
• Compliance: Always monitor the --early-stop-threshold to ensure you are not consuming excessive API credits during long-running optimization sessions. If the model is not responding, check your API key scope.