CtrlK
BlogDocsLog inGet started
Tessl Logo
SkillPublic

vector-and-embedding-weaknesses

Hunt vector / embedding weaknesses (OWASP LLM08:2025) — adversarial inputs against the RAG / similarity layer that cause cross-tenant leak, embedding-inversion privacy loss, semantic confusion, and retriever-driven prompt injection.

Quality

75%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Critical

Do not install without reviewing

Fix and improve this skill with Tessl

tessl review fix ./packages/decepticon/decepticon/skills/standard/analyst/vector-and-embedding-weaknesses/SKILL.md
SKILL.md
Quality
Evals
Security

Quality

Content

75%Weight 40%Scale 1-5

Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.

A dense, well-structured offensive-audit skill body: executable recon commands, concrete payloads, explicit finding-validation with negative controls, and ready CVSS vectors, all without tutorial padding. The main improvements are tightening the intro, making the Python PoC runnable, and expressing the audit steps as an explicit ordered sequence.

Suggestions

Make the embedding-inversion Python PoC runnable: import json, define the probe list concretely, and show how LEAKED_EMB and API are supplied.

Convert the audit workflow into an explicitly ordered step sequence (recon -> retrieve call sites -> check tenancy filters -> check ingest gating) so the sequence is stated, not implied by section numbers.

Trim the opening motivational paragraph to one or two sentences.

DimensionReasoningScore

Conciseness

The body is lean: it never explains what embeddings or vector stores are, and attack vectors are described in 2-4 lines each. The two-paragraph motivational intro ("The vector store is the silent half of an LLM application...") is the only trimming candidate — minor over-explanation, matching anchor 4 rather than anchor 5's every-token-earns-its-place.

4 / 5

Actionability

Concrete, copy-paste-ready grep commands for recon, explicit PoC payload strings ("XYZZY-{rand}", the refund injection text), and per-retrieval/per-indexer audit questions. The Python inversion PoC contains placeholders ("probes = [...candidate strings...]", undefined LEAKED_EMB/API, missing import json), which keeps it at anchor 4 (mostly executable, minor gaps) rather than 5.

4 / 5

Workflow Clarity

Sections 1-8 imply a coherent sequence (recognize -> audit -> exploit -> validate -> score -> chain), and the validate_finding contract provides explicit validation with negative controls ("same query from a clean account", "same query before the planted document is indexed"). It sits at anchor 4 rather than 5 because the audit workflow itself is a grep set plus checklists rather than an explicitly ordered step sequence.

4 / 5

Progressive Disclosure

A single-file skill (~156 lines) with no bundle files; content is organized into eight well-labeled sections and everything is appropriately inline at this size. It matches anchor 4 — good structure, minor organization gaps — since the PoC payload set or CVSS tables could be split into reference files if the skill grows.

4 / 5

Total

16

/

20

Passed

Description

66%Weight 40%Scale 1-5

Based on the skill's description, can an agent find and select it at the right time? Clear, specific descriptions lead to better discovery.

A specific, well-scoped description with strong domain language and concrete weakness classes. Its main gap is the absence of an explicit "Use when..." trigger clause inside the description field itself, which both caps completeness and slightly weakens natural triggering.

Suggestions

Append a trigger clause to the description, e.g. "Use when auditing RAG pipelines, vector stores, or retriever security, or when a user mentions embeddings, similarity search, or cross-tenant document leaks."

Add natural synonyms such as "vector store", "vector database", and common vendor names (Pinecone, pgvector, Weaviate) to broaden the terms users actually say.

Clarify the action scope beyond "Hunt" — e.g. "audit, exploit, and validate" — to round out the capability list.

DimensionReasoningScore

Specificity

The description names the domain ("vector / embedding weaknesses (OWASP LLM08:2025)", "RAG / similarity layer") and enumerates four concrete failure classes ("cross-tenant leak", "embedding-inversion privacy loss", "semantic confusion", "retriever-driven prompt injection"). It stops short of anchor 5 because it uses a single verb ("Hunt") rather than a comprehensive list of distinct actions, but exceeds anchor 3's 1-2 concrete actions.

4 / 5

Completeness

The "what" is clear and specific, but the description field itself contains no "Use when..." clause or equivalent trigger guidance — the trigger keywords live in metadata.when_to_use, outside the description. Per the judging guideline this caps completeness at 3, matching the anchor "clear 'what' but 'when' is missing".

3 / 5

Trigger Term Quality

Natural keywords a user would say are present: "vector", "embedding", "RAG", "cross-tenant leak", "prompt injection", "retriever", "similarity". A few natural terms are missing ("vector store", "vector database", vendor names like Pinecone/pgvector), so it fits anchor 4 rather than anchor 5's comprehensive synonym coverage.

4 / 5

Distinctiveness Conflict Risk

It carves a clear niche (the OWASP LLM08 vector/embedding layer), but "retriever-driven prompt injection" overlaps a general LLM01 prompt-injection skill, giving minor overlap risk — anchor 4 rather than anchor 5's minimal conflict risk.

4 / 5

Total

15

/

20

Passed

Validation

93%

Checks the skill against the spec for correct structure and formatting. All validation checks must pass before discovery and implementation can be scored.

Validation — 15 / 16 Passed

Validation for skill structure

CriteriaDescriptionResult

metadata_version

'metadata.version' is missing

Warning

Total

15

/

16

Passed

Repository

PurpleAILAB/Decepticon

Path

packages/decepticon/decepticon/skills/standard/analyst/vector-and-embedding-weaknesses/SKILL.md

Reviewed

Table of Contents

Is this your skill?

If you maintain this skill, you can claim it as your own. Once claimed, you can manage eval scenarios, bundle related skills, attach documentation or rules, and ensure cross-agent compatibility.