Feature Extractor

packages/core/src/triage/feature-extractor.ts exports a 45-element numeric vector used as a fast first-pass triage signal:

• extractFeatures(finding): number[]
• FEATURE_NAMES: string[]

The extractor is intentionally cheap:

• pure regex and string operations
• no LLM calls
• no network requests
• no external scanners

It is the handcrafted half of pwnkit’s VulnBERT-inspired hybrid direction.

Why it exists

VulnBERT’s reported ablations are the key reference point:

Model family	Recall	False positive rate
Handcrafted features alone	76.8%	15.9%
CodeBERT alone	84.3%	4.2%
Hybrid features + neural fusion	92.2%	1.2%

pwnkit uses the handcrafted layer today because it is deterministic, explainable, and cheap enough to run before any paid verification.

Group breakdown

Group	Count	Indices
Response features	13	0-12
Request features	10	13-22
Metadata features	8	23-30
Text quality features	10	31-40
Cross-field features	4	41-44

Full reference

Response features

Idx	Name	Type	Range	Rationale
0	resp_http_status	int	0+	Raw HTTP status extracted from evidence response
1	resp_sql_error	bool	0/1	SQL error strings are strong exploit evidence for SQLi
2	resp_stack_trace	bool	0/1	Stack traces often indicate unexpected execution paths
3	resp_error_message	bool	0/1	Generic error responses raise exploit plausibility
4	resp_payload_exact_reflection	bool	0/1	Exact reflection is strong for reflected XSS / echo-driven bugs
5	resp_payload_partial_reflection	bool	0/1	Partial reflection catches normalized or truncated echoes
6	resp_sensitive_data	bool	0/1	Leaked credentials / tokens / PII are high-signal evidence
7	resp_flag_pattern	bool	0/1	CTF-style flag markers are direct benchmark proof
8	resp_content_type_match	bool	0/1	Matching content type helps align evidence with the claimed category
9	resp_length	int	0+	Response size often distinguishes empty errors from real leakage
10	resp_waf_signature	bool	0/1	WAF blocks can explain false negatives and noisy probes
11	resp_redirect	bool	0/1	Redirect behavior matters for auth and route-confusion findings
12	resp_5xx_status	bool	0/1	Server errors are noisy but still useful exploit context

Request features

Idx	Name	Type	Range	Rationale
13	req_sql_syntax	bool	0/1	SQL syntax in the request aligns with SQLi claims
14	req_xss_payload	bool	0/1	Script / event-handler patterns align with XSS claims
15	req_ssti_syntax	bool	0/1	Template delimiters align with SSTI claims
16	req_path_traversal	bool	0/1	Traversal markers align with file-read claims
17	req_command_injection	bool	0/1	Shell metacharacters align with command-injection claims
18	req_encoding_detected	bool	0/1	Encoded payloads often show intentional bypass attempts
19	req_http_method	int	small ordinal	Request-method encoding lets the model distinguish GET/POST/etc.
20	req_auth_header	bool	0/1	Auth-bearing requests matter for IDOR / auth-boundary findings
21	req_param_count	int	0+	Parameter fanout helps characterize probe complexity
22	req_body_length	int	0+	Body size separates tiny probes from full exploit payloads

Metadata features

Idx	Name	Type	Range	Rationale
23	meta_severity_ordinal	int	0-4	Encodes severity as an ordinal prior
24	meta_confidence	float	0-1 typical	Carries the agent’s own confidence into triage
25	meta_high_confidence_category	bool	0/1	Some categories are more reliable than others in practice
26	meta_injection_class	bool	0/1	Injection bugs share structural traits worth flagging
27	meta_access_control_class	bool	0/1	Access-control bugs differ materially from injection bugs
28	meta_has_template_id	bool	0/1	Template-backed findings tend to be more structured
29	meta_has_cwe	bool	0/1	CWE references correlate with more mature analysis text
30	meta_has_cve	bool	0/1	CVE references often indicate external corroboration

Text quality features

Idx	Name	Type	Range	Rationale
31	text_description_length	int	0+	Very short descriptions often correlate with low-quality findings
32	text_repro_steps	bool	0/1	Reproduction steps are a strong signal of finding quality
33	text_impact_statement	bool	0/1	Explicit impact reasoning raises trust in the finding
34	text_hedging_language	bool	0/1	Hedging often correlates with weak or speculative findings
35	text_verification_language	bool	0/1	”confirmed”, “verified”, “reproduced” are strong positive cues
36	text_analysis_length	int	0+	Richer analysis text often means a more grounded claim
37	text_code_blocks	bool	0/1	Embedded code or PoC snippets raise exploit credibility
38	text_evidence_request_nonempty	bool	0/1	Missing request evidence is a major weakness
39	text_evidence_response_nonempty	bool	0/1	Missing response evidence is a major weakness
40	text_evidence_analysis_nonempty	bool	0/1	Missing analyst context weakens triage quality

Cross-field features

Idx	Name	Type	Range	Rationale
41	cross_payload_category_consistent	bool	0/1	Checks that the request payload matches the claimed bug class
42	cross_severity_confidence_interaction	float	0+	Multiplies severity prior by agent confidence
43	cross_response_request_length_ratio	float	0+	Large response / request ratios can indicate leakage or reflected amplification
44	cross_evidence_completeness	float	0-1	Non-empty request / response / analysis count divided by 3

How features are chosen

The features deliberately mix five signal families:

• response evidence
• request payload shape
• metadata priors
• text-quality heuristics
• cross-field consistency

That split reflects how triage actually works in practice: a finding is not credible because any one field looks good, but because several independent signals line up.

Fusion with neural models

The extractor is meant to pair with text embeddings, not replace them.

The natural hybrid shape is:

1. text -> encoder embedding
2. features -> linear projection
3. fusion head over both representations
4. binary TP/FP classifier

That is the same broad architecture family referenced from Finding Triage ML.

Domain-transfer caveat

These features are strongest on web and exploit-style findings because many of them look for:

• reflected payloads
• status codes
• stack traces
• SQL errors
• request / response evidence density

On npm supply-chain findings, many of those fields are sparse or zero. That is not a bug in the extractor; it is a real domain-transfer limitation and exactly the kind of thing a paper should report honestly.

Reproducibility

• deterministic
• source available
• pure local computation
• no model calls
• no network dependence

That makes the feature vector suitable for ablations, baselines, and air-gapped experimentation.

Related

• Triage Dataset — where this vector is emitted into JSONL
• Finding Triage ML — broader design and hybrid-model direction
• FP Reduction Moat — how the feature layer fits into the overall stack