ASL V6: Open-source red-teaming tool detects AI agent flaws with 98% fewer false alerts

Siva Aditya Panuganti, a security researcher with six published CVEs and GHSAs in production AI systems, has released ASL V6, a research-grade vulnerability assessment and red-teaming engine designed specifically for Python and AI agent codebases. The tool is free and open-source under the MIT License.

ASL V6 combines three complementary techniques to identify and verify security flaws. First, it runs 10 security analyzers that examine code for OWASP Top 10 LLM and agent vulnerabilities, including prompt injection sinks, goal hijacking, unsafe code execution (via eval/exec), and tool abuse. Second, it uses Abstract Syntax Tree (AST) contextual filtering to parse Python syntax trees and ignore test suites, mock files, and docstrings—this step alone eliminates around 98% of false positive alerts. Third, it performs live Docker runtime verification by running untrusted code snippets inside isolated ephemeral Docker containers (python:3.11-slim) to confirm whether an injection is actually exploitable in a live environment.

All execution happens locally on the user's machine; no data is sent over the internet. If users want LLM-assisted code patch suggestions, they can optionally provide a free NVIDIA developer API key via the NVIDIA_API_KEY environment variable. If no key is provided, the tool automatically uses offline, deterministic AST-based patch generation at zero cost. In testing on real repositories, ASL V6 achieved a 97.9% false positive reduction on LangGraph (606 raw alerts down to 13 validated true positives) and 100% on the ASL V6 engine itself (48 raw alerts down to 0)—both verified in runtime.

Users can install the tool via GitHub (https://github.com/sivaadityacoder/asl-v6.git), run a local security audit on their project directory, and execute benchmarks on multiple repositories. The package includes a GitHub Actions CI/CD workflow (asl_v6_ci_cd_action.yml) to automatically gate pull requests for security flaws. For teams that need deeper expertise, Panuganti offers paid advisory services: a monthly architecture review retainer at $2,500/month, emergency pre-launch or post-incident security assessments at $5,000 fixed, and EU AI Act technical documentation services at $3,500–$5,000 fixed. He is currently accepting 3 advisory clients for Q3 2026 and can be reached at adityasecuritylabs@gmail.com.

ASL V6 addresses a real pain point in AI agent security: the flood of false positives that comes from naive code scanning. By combining Abstract Syntax Tree (AST) parsing—which understands Python syntax at a deep level—with contextual filtering that ignores test files and docstrings, the tool eliminates around 98% of noise before it reaches developers. The addition of live Docker runtime verification is the crucial second step: it confirms that an alert represents an actual exploitable vulnerability, not a theoretical one. This two-stage approach reflects the author's background in responsible disclosure and CVE research; six published CVEs and GHSAs in production AI systems (AutoGPT, FlowiseAI, Trilium Notes, and LLaMmlein) give credibility to the threat model.

The decision to keep execution entirely local is significant for adoption. Teams building proprietary or regulated AI agents—especially in Europe, where the EU AI Act is now in effect—can run the scanner on sensitive code without routing it through a third party. The optional NVIDIA API integration for patch suggestions is a good compromise: it adds value for teams that want it without creating a hard dependency or requiring credentials upfront. The tool's integration with CI/CD pipelines via GitHub Actions suggests the author understands deployment workflows and is positioning ASL V6 as something that fits into existing development processes.

No comments yet. Be the first to share your thoughts!