Vector and Embedding Security Tester
Tests RAG vector stores for embedding inversion, tenant leakage, and poisoning, with authorization.
- Repository
- mukul975/Anthropic-Cybersecurity-Skills
- Stars
- 35,430
- Licence
- Apache-2.0
- Files
- 4 files
What it does
When installed, the agent gains a structured playbook for security-testing the retrieval layer of retrieval-augmented generation (RAG) apps, the part that stores documents as numeric embeddings and fetches them to answer questions. It guides the agent through checks for embedding inversion (recovering source text from vectors), membership inference, cross-tenant data leakage, knowledge-base poisoning, and hidden prompt-injection text in retrieved chunks. It also includes a Python helper script that runs four of these checks against a Qdrant test instance.
Great for
- Security engineers performing an authorized assessment of a RAG application
- Teams auditing whether a multi-tenant vector database properly isolates customer data before launch
Skip it if
- Testing systems you do not own or have no written authorization to assess, since inversion and cross-tenant probing can expose third-party data
- Running against production corpora, because poisoning tests write data and could contaminate a live knowledge base
Try asking
- “Check whether my Qdrant collection leaks other tenants' chunks when I filter by tenant_id”
- “Measure how well an attacker could reconstruct patient text from the embeddings my app stores”
- “Test whether untrusted uploaded documents can dominate retrieval with injected instructions”
Install in 3 steps
- 1
Download the skill
The ZIP contains the 4 original files, the Apache-2.0 licence and an attribution note.
assessing-vector-and-embedding-weaknesses.zip - 2
Add it to your agent
Claude Code (all projects)
mkdir -p ~/.claude/skills && unzip assessing-vector-and-embedding-weaknesses.zip -d ~/.claude/skills/Claude Code / Agent SDK (this project only)
mkdir -p .claude/skills && unzip assessing-vector-and-embedding-weaknesses.zip -d .claude/skills/Claude apps (web/desktop): Settings → Capabilities → Skills → Upload skill, and choose the ZIP.
- 3
Ask for the task
No special command needed — the agent reads the skill's description and loads it when your request matches. Try one of the prompts above.
Get better results
- Use only a dedicated test collection and get written scope before running any poisoning or isolation probe
- Review scripts/agent.py before running it, since it executes locally and connects to whatever URL you pass, and sentence-transformers may download model weights from the internet
What's inside
- references/api-reference.md
- references/standards.md
- scripts/agent.py
- SKILL.md
Redistributed unmodified from mukul975/Anthropic-Cybersecurity-Skills under the Apache-2.0 licence. The assessment was drafted with AI from the skill's files on 2026-10-29. Skills can instruct your agent to run code — read SKILL.md and any scripts before installing.