Z-Edgar/CoER-Defender

VISIONPricing:Input $0.431 / Cached $0.0862 / Output $1.12Concurrent Unit Cost:1Model Size:9BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 17, 2026Architecture:Transformer Featherless Exclusive Cold

CoER-Defender is a 9 billion parameter language model developed by Z-Edgar, based on the Qwen3.5-9B architecture, specifically designed to enhance the safety of tool-use tasks. It is trained to resist adaptive indirect prompt injection attacks, ensuring legitimate tool execution even under adversarial conditions. This model excels in maintaining task utility while significantly reducing attack success rates, making it suitable for secure AI agent deployments.

Loading preview...

CoER Defender: Robustness Against Prompt Injection

CoER-Defender is a 9 billion parameter model, built upon the Qwen3.5-9B architecture, specifically engineered to provide robust defense against adaptive indirect prompt injection in tool-use scenarios. Developed by Z-Edgar, this model is the final iteration of the CoER framework, refined through a multi-stage training process involving Attacker SFT, bilateral Co-PPO, and population-guided Defender SFT.

Key Capabilities

  • Advanced Prompt Injection Defense: Demonstrates high resilience against adaptive indirect prompt injection, with reported attack success rates as low as 0.25% on the main common-adaptive panel.
  • High Task Utility: Maintains strong task completion rates, achieving 75.40% Safe-U (task completion without compromise) on the main common-adaptive panel.
  • Secure Tool-Use: Designed to safely complete legitimate tool-use tasks even when subjected to sophisticated adversarial prompts.
  • Specialized Training: Benefits from a unique training regimen that includes supervised refinement on 5,760 trajectories, focusing on safe and successful task execution under attack.

Good For

  • Secure AI Agents: Ideal for deploying AI agents that interact with external tools and require high assurance against malicious inputs.
  • Research in AI Safety: Provides a strong baseline for further research into prompt injection defenses and adversarial robustness.
  • Applications Requiring Integrity: Suitable for systems where maintaining the integrity of tool execution and preventing unauthorized actions are critical.

Finetunes
1 models