Skip to content

Latest commit

 

History

1 Commit

Folders and files

NameName
Last commit message
Last commit date
 
 

Repository files navigation

AI/LLM Security Research

Independent security research on AI/LLM infrastructure — focusing on multi-tenant isolation, MCP protocol attack surface, and SSRF/cookie boundary leakage in widely-deployed LLM platforms.

10 fixes shipped across 5 major projects with a combined 335,000+ GitHub stars, including 5 GHSA advisory IDs (1 published, 4 reserved pending coordinated-disclosure window). 3 additional fixes currently in review (Dify, Microsoft autogen, crewAI).


GHSA Advisories

GHSA Project Class Fix PR Status
GHSA-2qwc-c2cc-2xwv Dify Tenant boundary violation dify#35797 reserved
GHSA-gvc6-fh3x-89xh Dify Path traversal dify#35796 reserved
GHSA-48xc-wmw8-3jr3 Dify Tenant boundary violation dify#35793 reserved
GHSA-2mr9-9r47-px2g Gradio Cross-Space cookie injection gradio#13384 reserved
GHSA-jmh7-g254-2cq9 Gradio SSRF (regression coverage) gradio#13388 published

By Project

Dify — 143k⭐

LLM application platform. Focus: tenant-scoping holes, plugin-daemon path traversal, key-rotation completeness.

PR Class Description
#35843 Key-rotation completeness reset-encrypt-key-pair left stale ciphertext in tool_builtin_providers / tool_api_providers / tool_mcp_providers — extended purge + regression tests
#35797 Tenant boundary FilePreviewApi text-extract endpoint accepted arbitrary tenant scope (GHSA-2qwc-c2cc-2xwv, reserved)
#35796 Path traversal Plugin-daemon file paths accepted ../ sequences (GHSA-gvc6-fh3x-89xh, reserved)
#35793 Tenant boundary App trace-config endpoints were not tenant-scoped (GHSA-48xc-wmw8-3jr3, reserved)
#34456 IDOR Data source binding missed tenant_id validation, enabling cross-tenant resource access
#36197 (in review) SSRF MCP server_url accepted internal addresses (127.0.0.1, link-local, RFC1918)

Cherry Studio — 46k⭐

Multi-provider LLM desktop client. Focus: MCP protocol attack surface in an Electron host.

PR Class Description
#13893 XSS dangerouslySetInnerHTML rendered untrusted MCP server descriptions and search results unsanitized
#13892 URL validation + path traversal shell.openExternal accepted file:// and arbitrary schemes; companion path-traversal guard
#14083 Credential hygiene OAuth tokens for deleted MCP servers persisted in disk-backed store

Gradio — 42k⭐

ML demo / Spaces framework. Focus: cross-origin isolation and SSRF in the proxy layer.

PR Class Description
#13384 Cross-origin cookies /proxy= shared a single cookie jar across Hugging Face Spaces, enabling cookie injection across tenants (GHSA-2mr9-9r47-px2g, reserved)
#13388 SSRF regression test Regression coverage for Blocks.from_config proxy_url SSRF guard (GHSA-jmh7-g254-2cq9)

autogen — 58k⭐ (in review)

Microsoft agent-orchestration framework.

PR Class Description
#7611 Code execution sandbox Opt-in sandbox flag for LocalCommandLineCodeExecutor (addresses #7462)

crewAI — 52k⭐ (in review)

Multi-agent framework.

PR Class Description
#5307 Code injection Replaced eval() in the calculator tool template with an AST-based safe evaluator

Focus Areas

  • Multi-tenant isolation in LLM platforms — IDOR, tenant-scoping gaps, cross-tenant resource enumeration (Dify boundary work)
  • MCP protocol attack surface — SSRF on server_url, untrusted server-supplied content rendered in clients, OAuth credential lifecycle (Dify + Cherry Studio + autogen)
  • Cross-origin cookie & session isolation — shared jars, proxy boundary leakage (Gradio)
  • Code execution boundaries in agent frameworkseval() removal, sandbox affordances (crewAI + autogen)

Methodology

  • Coordinated disclosure via GitHub Security Advisories where applicable (5 GHSA IDs assigned across the 10 shipped fixes; 1 currently published, 4 reserved during the upstream disclosure window). Remaining fixes shipped as direct upstream fix(security): PRs or correctness-adjacent hardening.
  • Every fix ships with regression tests in the same PR
  • Root cause + impact + repro included in each PR description

Maintained by @xr843. Also building fojin (佛津) — open-source Buddhist digital humanities platform.

About

Independent AI/LLM security research — 10 shipped fixes, 5 GHSA advisories; focus on multi-tenant isolation, MCP protocol attack surface, SSRF/cookie boundary leakage

Topics

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors