VelaOrg/vela-guard-32b

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

VelaOrg/vela-guard-32b is a 32 billion parameter chat model, fine-tuned from Qwen/Qwen3-32B by Vela, designed to act as a security layer for AI agents. It specializes in three critical checks: identifying secrets (private keys, seed phrases, API keys, wallet addresses), detecting prompt injections, and evaluating proposed tool calls for safety and alignment with user requests. With a context length of 32768 tokens, it provides robust pre-execution validation for AI agent actions, achieving 87% injection detection and 100% risky tool call detection in real-world benchmarks.

Loading preview...

Vela Guard 32B: AI Agent Security Layer

Vela Guard 32B is a specialized 32 billion parameter model, fine-tuned from Qwen/Qwen3-32B, designed to enhance the security and reliability of AI agents by performing critical pre-execution checks. Developed by Vela, this model acts as a robust safeguard, ensuring that AI agents operate safely and as intended.

Key Capabilities

  • Secrets Detection: Identifies and flags sensitive information such as private keys, seed phrases, API keys, and wallet addresses within any text input (prompts, logs, tool outputs).
  • Injection Detection: Determines if a tool output contains malicious instructions aimed at manipulating the AI agent, achieving 87% detection of attacks in real-world scenarios.
  • Tool Call Validation: Evaluates proposed tool calls against the user's original request, deciding whether to allow, confirm, or block the action. It caught 100% of risky calls in benchmarks.
  • High Accuracy: Achieves 0.908 injection accuracy and 0.872 tool call accuracy on real-world benchmarks, significantly outperforming smaller guard models.
  • JSON Output: Provides structured JSON responses for each check, making integration into automated workflows straightforward.

Good For

  • Securing AI Agents: Essential for applications where AI agents interact with sensitive data or perform actions that could have financial or security implications.
  • Preventing Prompt Injection: Mitigating risks associated with malicious user inputs attempting to hijack agent behavior.
  • Validating Tool Use: Ensuring that an agent's proposed actions (e.g., fund transfers, permission grants) align precisely with the user's intent.
  • Pre-execution Safety Checks: Implementing a crucial layer of defense before an agent processes information or executes a command.

Limitations

While strong in injection and tool call detection, secret recall is modest (61%), suggesting pairing with regex scanners for known key formats. The model is English-only and its training data is synthetic, though real-world benchmarks show strong performance.

Finetunes
1 models