Skip to content

USE-CASES.md

Latest commit

 

History

History
393 lines (295 loc) · 12.2 KB

USE-CASES.md

File metadata and controls

393 lines (295 loc) · 12.2 KB

SecID Use Cases

Concrete examples of what SecID enables. These aren't hypothetical - they're the problems we're solving.

Use Case 1: "What is this vulnerability, really?"

The Problem

You see CVE-2024-1234. You want to know:

  • What's the GHSA? (for package version info)
  • What's the CWE? (for weakness type)
  • What's the CVSS? (for severity)
  • Is there an exploit? (for prioritization)
  • What controls help? (for remediation)

Currently this requires:

  1. Search NVD → get CWE, CVSS (when enrichment available)
  2. Search GHSA → maybe find it, get package info
  3. Search OSV → maybe find aliases
  4. Search ATT&CK → different abstraction level
  5. Manually correlate everything

With SecID

secid:advisory/mitre.org/cve#CVE-2024-1234
    ├── aliases
    │   ├── secid:advisory/github.com/advisories/ghsa#GHSA-xxxx-yyyy-zzzz
    │   └── secid:advisory/google.com/osv#PYSEC-2024-567
    ├── classified_as
    │   └── secid:weakness/mitre.org/cwe#CWE-89
    ├── severity
    │   ├── cvss:3.1 → 8.8 (from NVD)
    │   └── epss → 0.42 (42% exploitation probability)
    ├── affects
    │   └── pkg:pypi/sqlparse@0.4.0-0.4.3
    ├── exploited_by
    │   └── secid:ttp/mitre.org/attack#T1190
    └── mitigated_by
        └── secid:control/owasp.org/asvs#V5.3.4

One query, complete picture.

Use Case 2: "Show me all prompt injection vulnerabilities"

The Problem

You're assessing LLM security. You want every known prompt injection issue.

But prompt injection spans:

  • CVEs (when assigned, which is rare)
  • GHSA (for package-level issues in LangChain, etc.)
  • Research papers (most prompt injection is academic)
  • ATLAS techniques (AML.T0043)
  • OWASP LLM Top 10 (LLM01)

No single database has them all. No standard way to query across.

With SecID

# Find everything classified as prompt injection
secid:weakness/mitre.org/cwe#CWE-1427  # Direct CWE for prompt injection
    └── instances
        ├── secid:advisory/mitre.org/cve#CVE-2023-29374  # LangChain
        ├── secid:advisory/mitre.org/cve#CVE-2023-36189  # LangChain
        ├── secid:advisory/github.com/advisories/ghsa#GHSA-...       # More LangChain
        └── ...

secid:weakness/owasp.org/llm-top10#LLM01  # OWASP category
    └── instances
        ├── (everything above)
        └── (additional research/incidents)

secid:ttp/mitre.org/atlas#AML.T0043  # Attack technique
    └── examples
        ├── (case studies)
        └── (related vulns)

Query any entry point, traverse to find everything related.

Use Case 3: "Map our controls to OWASP LLM Top 10"

The Problem

Auditor asks: "Show me how your controls address OWASP LLM Top 10."

You have controls documented somewhere. OWASP LLM Top 10 is documented. But the mapping between them is:

  • In someone's head
  • In a one-off spreadsheet
  • Outdated
  • Not queryable

With SecID

# Pre-built mapping in relationships/
mapping: "CSA AI Controls Matrix → OWASP LLM Top 10"

secid:weakness/owasp.org/llm-top10#LLM01  # Prompt Injection
    └── mitigated_by
        ├── secid:control/cloudsecurityalliance.org/aicm#INP-01  # Input validation
        │   └── strength: partial
        │   └── notes: "Helps but doesn't fully prevent"
        ├── secid:control/cloudsecurityalliance.org/aicm#INP-02  # Input sanitization
        │   └── strength: partial
        └── secid:control/cloudsecurityalliance.org/aicm#ARC-03  # Prompt isolation
            └── strength: strong

secid:weakness/owasp.org/llm-top10#LLM02  # Sensitive Information Disclosure
    └── mitigated_by
        ├── secid:control/cloudsecurityalliance.org/aicm#DAT-01  # Data classification
        ├── secid:control/cloudsecurityalliance.org/aicm#DAT-04  # Output filtering
        └── secid:control/cloudsecurityalliance.org/aicm#MOD-02  # Fine-tuning controls

Auditor can see exact mappings with rationale. You can generate compliance matrices automatically.

Use Case 4: "What's the AI security landscape?"

The Problem

New to AI security. Want to understand:

  • What databases track AI vulnerabilities?
  • What frameworks exist for AI threats?
  • What standards apply?
  • Who are the key players?

Currently: Read a bunch of blog posts, hope you find the right ones.

With SecID

Browse the entity registry:

secid:entity/mitre.org/atlas         # AI attack framework
secid:entity/owasp.org/llm-top10     # AI risk categories
secid:entity/cloudsecurityalliance.org/aicm            # AI controls
secid:entity/nist.gov/ai-rmf         # AI risk management

secid:entity/openai.com/gpt-4        # GPT-4 model
secid:entity/openai.com/chatgpt      # ChatGPT service
secid:entity/anthropic.com/claude    # Claude model
secid:entity/google.com/gemini       # Gemini model

# Each entity has:
# - What it is
# - What it contains
# - How to access it
# - Common misconceptions
# - Relationships to other entities

The entity files ARE the learning resource. Reading about ATLAS teaches you what it is, why it matters, and how it relates to everything else.

Use Case 5: "Track AI security issues that need rapid coordination"

The Problem

You discover a jailbreak technique that affects all LLMs. Traditional vulnerability tracking is challenging because:

  • No specific software version affected (affects model classes)
  • Behavioral pattern rather than code defect
  • Rapid disclosure pace in AI research community

Researchers need a way to track and coordinate on these issues while the broader ecosystem adapts.

With SecID (Future AI Vuln DB)

secid:advisory/ai-vuln-db/JB-2024-0042
    ├── name: "DAN 15.0 Jailbreak"
    ├── description: "Persona-based jailbreak effective against GPT-4, Claude, Gemini"
    ├── classified_as
    │   ├── secid:weakness/owasp.org/llm-top10#LLM01
    │   └── secid:ttp/mitre.org/atlas#AML.T0051
    ├── affects
    │   ├── secid:entity/openai.com/gpt-4
    │   ├── secid:entity/anthropic.com/claude-3
    │   └── secid:entity/google.com/gemini
    ├── discovered: 2024-03-15
    ├── references
    │   ├── https://jailbreakchat.com/...
    │   └── https://arxiv.org/...
    └── status: "unpatched"

Tracked, classified, queryable - complementing CVE for issues that benefit from faster coordination.

Use Case 6: "Get severity data from multiple sources"

The Problem

CVE-2024-9999 published. NVD enrichment is pending. You need:

  • CVSS score for prioritization
  • CWE for understanding
  • Affected versions for patching

Meanwhile, the vendor has published their own advisory with this information.

With SecID Overlays

# overlays/enrichment/CVE-2024-9999.yaml
overlay_type: enrich
target: secid:advisory/mitre.org/cve#CVE-2024-9999
source: "secid:advisory/acme/SA-2024-042"

adds:
  cvss_v3_1:
    score: 9.8
    vector: "CVSS:3.1/AV:N/AC:L/PR:N/UI:N/S:U/C:H/I:H/A:H"
    source: "vendor"

  cwe: "CWE-502"

  affected_versions:
    - ">=1.0.0 <1.5.3"
    - ">=2.0.0 <2.1.1"

  fixed_versions:
    - "1.5.3"
    - "2.1.1"

rationale: "Supplementing with vendor advisory data"
confidence: "high"

Get actionable data immediately from vendor sources while official enrichment completes.

Use Case 7: "Cross-reference everything for this package"

The Problem

You use langchain. You want EVERY security issue:

  • CVEs
  • GHSAs
  • Security advisories
  • Related vulnerabilities in dependencies
  • Relevant attack techniques

With SecID

pkg:pypi/langchain
    └── vulnerabilities
        ├── secid:advisory/mitre.org/cve#CVE-2023-29374
        │   └── classified_as: secid:weakness/mitre.org/cwe#CWE-94
        ├── secid:advisory/mitre.org/cve#CVE-2023-36189
        │   └── classified_as: secid:weakness/mitre.org/cwe#CWE-1427
        ├── secid:advisory/github.com/advisories/ghsa#GHSA-...
        └── secid:advisory/github.com/advisories/ghsa#GHSA-...

    └── dependencies (from SBOM)
        ├── pkg:pypi/openai
        │   └── vulnerabilities: [...]
        ├── pkg:pypi/tiktoken
        │   └── vulnerabilities: [...]
        └── ...

    └── relevant_techniques
        ├── secid:ttp/mitre.org/atlas#AML.T0043  # Prompt injection
        ├── secid:ttp/mitre.org/atlas#AML.T0054  # Plugin compromise
        └── ...

Complete security picture for dependency decisions.

Use Case 8: "Understand coverage across vulnerability databases"

The Problem

You want to understand which vulnerability types are well-tracked across the ecosystem and where emerging areas might need attention.

With SecID Analysis

# Query: Compare coverage across databases by weakness type

secid:weakness/mitre.org/cwe#CWE-1427  # Prompt Injection
    └── cve_count: 12
    └── ghsa_count: 45
    └── atlas_references: 3
    └── coverage_note: "Emerging area, coverage expanding"

secid:weakness/mitre.org/cwe#CWE-1426  # Improper AI Output Validation
    └── cve_count: 3
    └── ghsa_count: 18
    └── coverage_note: "New weakness type, tracking growing"

secid:weakness/mitre.org/cwe#CWE-502   # Deserialization
    └── cve_count: 2,847
    └── ghsa_count: 312
    └── coverage_note: "Mature, well-tracked"

Data-driven insights into ecosystem coverage by vulnerability type.

Use Case 9: "Generate a threat model"

The Problem

Building an LLM application. Need threat model covering:

  • Relevant attack techniques
  • Applicable weaknesses
  • Recommended controls
  • Known vulnerabilities in your stack

With SecID

# Input: I'm building with LangChain + OpenAI + Pinecone

# Output threat model:

Components:
  - pkg:pypi/langchain → 15 known vulns
  - secid:entity/openai.com/api → prompt injection exposure
  - secid:entity/pinecone.io/vector-db → embedding manipulation risk

Relevant Threats:
  - secid:ttp/mitre.org/atlas#AML.T0043 (Prompt Injection)
      └── your_exposure: HIGH (direct user input to LLM)
  - secid:ttp/mitre.org/atlas#AML.T0054 (Plugin Compromise)
      └── your_exposure: MEDIUM (using tools)
  - secid:ttp/mitre.org/atlas#AML.T0048 (Embedding Manipulation)
      └── your_exposure: MEDIUM (RAG with Pinecone)

Applicable Weaknesses:
  - secid:weakness/owasp.org/llm-top10#LLM01 through LLM10

Recommended Controls:
  - secid:control/cloudsecurityalliance.org/aicm#INP-01 (Input Validation)
  - secid:control/cloudsecurityalliance.org/aicm#OUT-01 (Output Validation)
  - secid:control/cloudsecurityalliance.org/aicm#RAG-01 (RAG Security)

Threat modeling becomes a query, not a research project.

Use Case 10: "Keep AI security research organized"

The Problem

AI security research is scattered:

  • Papers on arXiv
  • Jailbreaks on GitHub/Reddit
  • Vendor disclosures
  • Conference talks
  • Blog posts

No central index. No way to track what's been tried.

With SecID

# Research paper example (using reference type)
secid:reference/arxiv.org/1908.07125
    ├── title: "Universal Adversarial Triggers"
    ├── demonstrates
    │   └── secid:ttp/mitre.org/atlas#AML.T0043
    ├── affects
    │   ├── secid:entity/openai.com/gpt-2
    │   └── secid:entity/google.com/bert
    └── related_to
        └── secid:weakness/mitre.org/cwe#CWE-1427

# Jailbreak technique (future: could be tracked via AI vuln database)
secid:advisory/ai-vuln-db/JB-2024-0015
    ├── name: "DAN Jailbreak"
    ├── versions: ["DAN 1.0", "DAN 5.0", "DAN 11.0", ...]
    ├── demonstrates
    │   └── secid:ttp/mitre.org/atlas#AML.T0051
    └── affects
        ├── secid:entity/openai.com/gpt-3.5
        └── secid:entity/openai.com/gpt-4 (partial)

Research becomes navigable, not just searchable.

Common Thread

All these use cases share the same pattern:

  1. Single identifier format across all security knowledge
  2. Relationships that connect disparate databases
  3. Overlays that fill gaps without modifying sources
  4. Entity registry that defines what exists

SecID doesn't replace existing databases. It makes them work together.