SuperClaw – Open-Source Framework to Red-Team AI Agents for Security Testing

Superagentic AI has introduced SuperClaw, an open-source framework designed for pre-deployment security testing of autonomous AI coding agents.

Launched in late 2025, SuperClaw addresses a significant issue in enterprise AI integration, where agents are often deployed with extensive tool access and high privileges without comprehensive security validation prior to deployment.

The primary focus of SuperClaw's development is to test how an autonomous AI agent behaves under adversarial conditions, rather than just assessing its configuration. This is crucial as traditional security scanners are not equipped to handle the dynamic and adaptive nature of these agents.

SuperClaw conducts scenario-driven, behavior-first security evaluations on real agents within controlled environments.

It utilizes its Bloom scenario engine to generate adversarial scenarios, executing them against live or mock agent targets. The framework records comprehensive evidence, including tool calls and output artifacts, and evaluates results against explicit behavior contracts. These contracts define the intent, success criteria, and mitigation guidance for each security property.

The framework supports five primary attack techniques:

Prompt Injection: Direct and indirect methods. Encoding Obfuscation: Includes Base64, hex, Unicode, and typoglycemia. Jailbreaks: Techniques like DAN, role-play, and grandmother bypasses. Tool-Policy Bypass: Exploits alias confusion. Multi-Turn Escalation: Across conversation turns.

SuperClaw evaluates security behaviors that span critical risks like prompt-injection resistance and sandbox isolation, high-severity concerns such as tool-policy enforcement and cross-session boundary integrity, and medium-severity issues like configuration drift detection and ACP protocol security.

Attack Technique Description Agent Testing Focus

Prompt Injection Malicious prompts attempt to override system instructions and hijack decision-making. Tests if the agent can detect and reject injected instructions.

Encoding Utilizes Base64, hex, Unicode, or typoglycemia to obscure malicious intent. Evaluates if the agent can spot and refuse encoded payloads.

Jailbreak Bypasses guardrails using techniques like DAN prompts and role-play. Assesses the agent's resilience to safety bypass attempts.

Tool-Bypass Exploits tool aliases or weak policies to misuse tools. Checks if the agent adheres to strict allow/deny rules for tools.

Multi-Turn Builds from benign queries to malicious objectives over several turns. Evaluates how the agent manages long-context interactions.

Reports are available in HTML for review, JSON for automation, or SARIF for integration with GitHub Code Scanning and CI/CD workflows.

SuperClaw integrates with CodeOptiX, Superagentic AI's code evaluation engine, offering combined security and optimization assessments within a single pipeline.

SuperClaw operates with strict built-in guardrails. By default, it functions in local-only mode, blocking remote targets to prevent unauthorized use. Remote connections require a valid SUPERCLAW_AUTH_TOKEN password from the target system's administrator.

The project mandates written authorization before testing and emphasizes that automated findings are indicators for manual verification, not definitive proof of exploitation.

SuperClaw is accessible on GitHub under the Apache 2.0 license and can be installed via pip install superclaw . It is part of the Superagentic AI ecosystem, alongside SuperQE and CodeOptiX, aimed at development teams requiring robust agent security pre-deployment.

Based on reporting by Cyber Security News.