Skip to content

adversa-ai/research

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

32 Commits
 
 
 
 

Repository files navigation

Adversa AI Research

Public attachments for Adversa AI security research and disclosures. Each subdirectory under artifacts/ corresponds to a single published advisory or article and contains the full technical report, proof-of-concept code, and supporting materials.

This repository is intended for security practitioners who want the technical detail behind a public Adversa AI write-up — readers who only need the narrative version should start with the linked blog post inside each artifact.

Layout

artifacts/
└── <article-slug>/
    ├── README.md      # Short index: what this advisory is and what's in the directory
    ├── report/        # Full technical report and screenshots
    └── poc/           # Safe, reproducible proof-of-concept (where applicable)

Published research

Artifact Topic Blog post
artifacts/trustfall-mcp-settings-rce/ Project-scoped MCP auto-approval enables 1-click RCE across four agentic CLIs (Claude Code, Gemini CLI, Cursor CLI, Copilot CLI); 0-click in Claude Code CI via claude-code-action. Same project-scoped settings convention as CVE-2025-59536; classified by Anthropic as design intent. TrustFall: One-keypress RCE in agentic coding CLIs via project-scoped MCP settings

Disclosure posture

All findings published here have been reported to the affected vendor first. Where a vendor has declined a finding as out-of-scope of their threat model, the corresponding write-up makes the vendor's position explicit and documents the residual risk for defenders.

Proof-of-concept code is deliberately scoped to the minimum that demonstrates the vulnerability. PoCs intended for general execution on a developer machine do not exfiltrate data, open network connections, or modify anything outside the working directory. A small number of PoCs are scoped to self-testing on infrastructure the reader owns — for example, a CI fixture that requires the reader to point an exfiltration target at their own collector URL on a repository they control; nothing leaves the runner until the reader supplies that URL. These are flagged in their per-artifact README and require explicit setup steps. Read the per-artifact README before running anything.

Contact

Security research and disclosure inquiries: https://adversa.ai/

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

 
 
 

Contributors