How We Score Articles
Every article published on Grid the Grey is evaluated by the Claude AI model across two dimensions: Relevance Score and Threat Level. This ensures only the most significant AI security developments reach you.
Relevance Score (0–10)
The relevance score measures how closely an article relates to AI security, machine learning threats, LLM vulnerabilities, or related adversarial techniques.
| Score Range | Meaning |
|---|---|
| 8.0 – 10.0 | Highly relevant — direct AI/ML security impact, novel techniques, or active exploitation |
| 6.0 – 7.9 | Relevant — significant AI security implications, important for practitioners |
| Below 6.0 | Not published — insufficient AI security relevance |
Factors that raise the score:
- Direct exploitation of AI/ML systems
- Novel attack techniques against LLMs or ML pipelines
- Active CVEs affecting AI infrastructure
- Research from high-credibility sources (Google Project Zero, CISA, academic institutions)
- Techniques mappable to MITRE ATLAS or OWASP LLM Top 10
Factors that lower the score:
- Generic cybersecurity news with weak AI/ML connection
- Marketing content or vendor announcements without technical depth
- Duplicated coverage of an already-published story
Threat Level
The threat level classifies the operational urgency of a security issue:
| Level | Colour | Meaning |
|---|---|---|
| CRITICAL | Red | Active exploitation in the wild. Immediate review and action required. |
| HIGH | Orange | High probability of exploitation. Prioritise patching or mitigation. |
| MEDIUM | Yellow | Moderate risk. Monitor for developments and plan remediation. |
| LOW | Green | Limited impact. Standard review cycle is appropriate. |
Threat level is assessed based on: CVSS scores where available, evidence of active exploitation, breadth of affected systems, and the sophistication of the attack technique.
Framework Mapping
In addition to scoring, every article is mapped (where applicable) to:
- MITRE ATLAS — adversarial ML attack technique IDs (e.g. AML.T0051)
- OWASP LLM Top 10 — LLM vulnerability categories (e.g. LLM01)
This mapping is performed by the Claude AI model using official MITRE ATLAS and OWASP LLM Top 10 taxonomies, cross-referenced against article content.