KittyProxy and Autonomous Agent

This is where KittySploit stands apart from ``Metasploit-style'' frameworks: the intelligent proxy and local agent turn manual browsing into a structured attack plan. This chapter sets the end-to-end scenario; chapter 21 (Agent policies, local LLM, and campaign governance) details each --approve-risk option, TOML policy files, and Ollama integration.

Agent scenario schedule (2 hours)

KittyProxy in CLI mode

proxy start --host 127.0.0.1 --port 8080
proxy status

Configure your browser or:

export http_proxy=http://127.0.0.1:8080
export https_proxy=http://127.0.0.1:8080

Browse DVWA: login, SQLi, XSS, upload, Security level. KittyProxy captures HTTP/HTTPS flows, endpoints, parameters, technology hints, and risk signals.

proxy interactive --auto-start

Interactive mode lets you inspect requests without a dedicated web UI (see also the relevant documentation for in-depth coverage).

Proxy capture checklist

• POST login request (PHPSESSID, security cookies).
• At least one GET with injectable parameter (?id=, ?name=).
• One multipart upload.
• CSRF or password change form.

Agent --- reconnaissance only

Start without exploitation to observe behavior:

workspace use labcorp-dvwa-recon
agent doctor
agent http://127.0.0.1 --no-exploit --goal recon --safety-profile discreet

The agent typically chains: module scan, analysis, reasoning (local LLM if --llm-local, the relevant documentation), report.

Cautious modes

agent http://127.0.0.1 --no-exploit --plan-only --verbose
agent http://127.0.0.1 --profile safe-web --dry-run

• --dry-run : plan without network traffic.
• --plan-only : recon then stop before exploit.
• --safety-profile discreet : tightened budgets.

Keep the displayed run_id (e.g.\ agent_20260626T140000_abc123).

Agent enriched by proxy

After browsing via KittyProxy:

agent http://127.0.0.1 \
  --no-exploit \
  --reuse-proxy-auth \
  --proxy-flow-limit 40

The agent imports KittyProxy flows: cookies, endpoints, login forms, HTTP replay candidates. Workflow: I browse, the tool reasons intelligently.

Additional options (detail in the relevant documentation):

agent http://127.0.0.1 \
  --no-exploit \
  --reuse-proxy-auth \
  --http-replay safe \
  --http-replay-max 5

Validation with approvals

To go as far as active validation in the lab:

agent http://127.0.0.1 \
  --goal validate \
  --approve-risk active \
  --safety-profile normal \
  --max-modules 20 \
  --threads 5

Risk levels (increasing): read, active, intrusive, destructive. Each higher level includes the previous ones.

Mission profiles

agent profiles

Common profiles:

• safe-web : cautious recon, low budgets.
• internal-lab : more aggressive in Docker lab.
• Custom TOML profiles: the relevant documentation, policies section.

Adapt --request-budget and --max-modules to the mandate.

Resume and explainability

Agent campaigns are traced:

agent explain <run_id>
agent replay <run_id>
agent retest <finding_id>
agent --resume <run_id>
agent metadata <run_id>

• explain : module justification --- essential for client deliverable.
• replay : replay campaign for regression or demo.
• retest : after DVWA Medium fix.
• metadata : JSON export for CI/CD pipeline.

HTTP replay

agent http://127.0.0.1 \
  --http-replay safe \
  --http-replay-max 3 \
  --reuse-proxy-auth

active mode (mutations) only with --approve-active-replay and explicit ROE agreement.

Complete end-to-end scenario

1. lab start dvwa-basics
2. workspace use labcorp-dvwa-recon
3. agent doctor
4. proxy start + manual DVWA browsing (15--20 min)
5. agent http://127.0.0.1 --no-exploit --reuse-proxy-auth
6. agent explain <run_id> : compare with manual findings ch. 7--8
7. Manual validation of gaps (agent false positives / negatives)
8. agent ... --goal validate --approve-risk active --max-modules 15
9. campaign --preview
10. campaign --formats report,navigator

Agent vs manual workflow comparison

Quick troubleshooting

agent doctor
doctor
proxy status

• Agent with no results: verify DVWA up, proxy cookies, scope.
• Local LLM absent: campaign without enriched reasoning.
• Too many requests: lower --threads, --request-budget.

Toward the relevant documentation

Before a client engagement with the agent, read the relevant documentation in full:

• define a TOML policy file (allowed_modules, deny_destructive);
• map --safety-profile to contractual requirements;
• configure --llm-local and secrets;
• integrate retest into the post-fix quality loop.

the relevant documentation had you practice; the relevant documentation teaches you to govern.