JevSec

Self-hosted behavioral security triage powered by local Qwen3-4B.

Your WAF sees requests. JevSec sees behavior.

Project page · Technical write-up · Benchmark · Architecture · Roadmap

Why JevSec exists

Request-level WAFs are strong at matching exploit syntax and known request patterns. JevSec is aimed at the layer above that: short sequences of activity that become suspicious only when viewed together.

JevSec groups web activity into behavior windows, combines deterministic evidence with local-model decisions, and produces structured findings for human review.

Nginx / JSONL
      │
      ▼
privacy-aware normalization
      │
      ▼
behavior windows
      │
      ├── deterministic rules
      │
      └── local Qwen3-4B
               │
               ▼
BENIGN / REVIEW / HIGH_RISK / UNCERTAIN
               │
               ▼
       local SQLite + dashboard

JevSec complements a WAF. It does not replace WAF enforcement.

Benchmark

Semi-real CSIC replay

Deterministic held-out sample:

• 250 test windows
• 145 anomalous / 105 normal
• local Qwen3-4B
• thresholds selected from validation only
• shadow-mode review, not automated blocking

System	Recall	Precision	Observed FPR	Anomalies detected
Static rules	20.69%	100.00%	0.00%	30 / 145
Qwen3-4B	23.45%	97.14%	0.95%	34 / 145
JevSec Hybrid	26.90%	97.50%	0.95%	39 / 145

Headline: Hybrid detected 39 vs 30 anomalous windows compared with the static-rule baseline — 30% more detections — while adding 1 false review among 105 normal windows.

For windows containing only one anomalous request mixed with normal requests, recall moved from 16.67% to 25.00% (+50% relative on that small subgroup).

What the benchmark does not prove

Current risk-score AUROC remains weak:

• Static rules: 0.522
• Qwen3-4B: 0.473
• Hybrid: 0.454

So the result is evidence of incremental review coverage, not evidence that the model globally ranks risk well, detects unknown vulnerabilities in production, or outperforms a mature WAF at request-level exploit signatures.

See public benchmark v2, OWASP CRS comparison, and failure analysis.

JevSec vs a traditional WAF

	Traditional request-level WAF	JevSec
Primary view	Individual HTTP request	Activity across a behavior window
Best fit	Request syntax / signatures / exploit indicators	Sequence, frequency, context, rule-model disagreement
Output	Allow / block / anomaly score	Review-oriented risk finding
Model	Usually not required	Local Qwen3-4B
Deployment	Enforcement layer	Shadow / triage layer
Relationship	Keep it	Add JevSec beside it

The OWASP CRS report in this repository intentionally treats the two systems as complementary rather than as interchangeable products.

Key properties

• Self-hosted — core application and decision workflow stay local.
• Local AI — current supported model is Qwen3-4B through local-jev.
• Privacy-aware context — the model receives whitelisted aggregate features rather than raw request paths, cookies, Authorization values, or user-agent strings.
• Behavior windows — aggregate by source IP and optional pseudonymous session identity.
• Explainable evidence — rule/model evidence is kept with each finding.
• Shadow mode first — no automatic firewall changes, bans, or request blocking.
• Reproducible evaluation — fixed seeds, validation-only threshold fitting, held-out test reporting.

Quick start

Requirements: Python 3.12, uv, Git.

git clone https://github.com/ccjmcc/jevsec.git
cd jevsec
./scripts/demo.sh

Open:

http://127.0.0.1:8000

The first local-model launch downloads Qwen3-4B weights.

UI-only smoke test:

SDE_DECISION_PROVIDER=mock ./scripts/demo.sh

Shadow mode

security-engine shadow --nginx /var/log/nginx/access.log

Shadow mode tails logs and emits structured findings. It does not block requests, change network configuration, or ban addresses.

Docker

On Linux:

cp .env.example .env
docker compose up --build

On Apple Silicon, run local-jev natively if you want Apple GPU acceleration.

Privacy and security

JevSec is intentionally conservative about what reaches the model:

• cookies are not retained;
• Authorization values are not retained;
• passwords and API keys are discarded;
• unknown JSON fields are discarded;
• user/session identifiers are pseudonymized before aggregation;
• model outputs are advisory.

Read PRIVACY.md, SECURITY.md, and THREAT_MODEL.md.

Project status

Research Alpha / shadow-mode engineering prototype.

The next milestones are real-world shadow validation, better sequence context, stronger risk ranking, and tighter WAF-signal fusion. See the roadmap.

Documentation

• Architecture
• Usage
• Benchmark index
• OWASP CRS comparison
• Failure analysis
• Privacy
• Security policy
• Threat model
• Press kit
• Roadmap
• Contributing

License

MIT. See LICENSE.