vuln-agent

Receives DefectDojo scan webhooks, analyzes High/Critical findings with Claude Haiku, and sends enriched security reports to a configurable webhook endpoint.

DefectDojo webhook (scan added)
    → POST /webhook
    → GET /api/v2/findings/?test=<id>&active=true&duplicate=false&severity=Critical,High
    → Claude Haiku — exploitability analysis
    → POST NOTIFY_URL (Slack attachment format)

Configuration

Required

Variable	Description
DEFECTDOJO_URL	DefectDojo base URL
DEFECTDOJO_API_KEY	DefectDojo API token
ANTHROPIC_API_KEY	Anthropic API key
NOTIFY_URL	Webhook destination for reports

Optional

Variable	Default	Description
DEFECTDOJO_PUBLIC_URL	DEFECTDOJO_URL	Public-facing URL used for finding links in notifications. Set when DEFECTDOJO_URL is an internal address
DATABASE_URL	—	PostgreSQL DSN. Enables search_infrastructure tool — Claude can query infra context before assessing risk
EXTRA_PROMPT_FILE	—	Path to a file appended to the system prompt at startup
NOTIFY_FORMAT	slack	slack (attachments) or text
MAX_FINDINGS	10	Maximum findings per notification, Critical first
LLM_CONCURRENCY	4	Maximum parallel Claude requests per webhook
IGNORE_SCAN_TYPES	—	Comma-separated scan type names to skip entirely, e.g. Gitleaks Scan,DAST Scan
LOG_LEVEL	INFO	Log level

Webhook

DefectDojo must be configured to POST scan notifications to /webhook.

Expected payload:

{
  "description": "...",
  "url_ui": "https://dojo.example.com/test/69"
}

test_id is extracted from url_ui via /test/(\d+). The endpoint returns 200 immediately; processing happens in the background.

Service identification

For each finding, the service name is resolved in this order:

1. finding.service — populated when the scan was imported with a service parameter
2. First path component of finding.file_path — e.g. payment-service/src/Foo.java → payment-service
3. No service — AI analysis is skipped for this finding (sent as "Analysis unavailable")

Infrastructure search (optional)

When DATABASE_URL is set, Claude receives the search_infrastructure tool and calls it before assessing risk. The tool does a text search against a infrastructure table:

SELECT kind, name, namespace, cluster, content
FROM infrastructure
WHERE name ILIKE '%<service>%' OR content ILIKE '%<service>%'
ORDER BY kind  -- Deployment, Ingress, HTTPRoute, Service first
LIMIT 5

The table is expected to contain Kubernetes resource records with a content text column describing each resource. The tool is intentionally simple — no embeddings or vector search required.

Prompt extension

The built-in system prompt covers the generic analysis task only. Environment-specific context (infrastructure topology, service naming conventions, custom rules) is injected at deploy time:

EXTRA_PROMPT_FILE=/app/config/extra_prompt.txt

The file content is appended to the system prompt at startup. Mount it from a ConfigMap, secret, or volume.

Notification format

Reports use a Slack-compatible payload compatible with Slack, RocketChat, and similar incoming webhook endpoints.

One message per scan, one attachment per finding:

my-repo | MR-42 | Trivy Scan | 14 findings (3 shown)

[Critical] CVE-2024-1234 — log4j-core 2.14.0
  Risk: High — public ingress detected, RCE via JNDI lookup is feasible without authentication.
  Remediation: upgrade log4j-core to 2.17.1+
  Service: payment-service | Component: log4j-core 2.14.0 | CVE: CVE-2024-1234

[High] ...

Set NOTIFY_FORMAT=text for plain-text output.

Running locally

export DEFECTDOJO_URL=https://dojo.example.com
export DEFECTDOJO_API_KEY=...
export ANTHROPIC_API_KEY=...
export NOTIFY_URL=http://localhost:8081/webhook

pip install -r requirements.txt
python main.py

Building

docker build -t vuln-agent:latest .