4 Passive reconnaissance agents

This text introduces reconnaissance as the foundation of offensive security work and argues that repeatable pipelines are safer and more useful than scattered manual commands or one-off scripts. Instead of relying on ad hoc notes, screenshots, and terminal output, an AI-assisted reconnaissance workflow breaks the process into explicit stages: planning what to examine, acting through low-impact collection steps, observing results, and recording evidence. The key idea is that AI should augment human analysts rather than replace them, helping propose actions, interpret observations, and summarize findings while humans enforce boundaries, approve execution, and remain accountable.

The chapter walks through building a minimal Python-based reconnaissance pipeline that generates candidate hostnames, resolves DNS, sends lightweight HTTP HEAD requests, optionally inspects TLS metadata, checks for simple WAF or CDN hints, and records each observation as a structured JSONL artifact. The design emphasizes composability: each function performs a small stateless task, and each result is appended to a durable recon.jsonl file with fields such as host, IP addresses, headers, TLS details, notes, and timestamps. These artifacts become persistent memory for both humans and future AI agents, making it possible to replay, audit, search, cluster, and reason over reconnaissance activity as a chronological story rather than a pile of unstructured logs.

The text also stresses that automation must be constrained by scope, safety gates, and responsible data handling. Reconnaissance pipelines should enforce approved targets, time windows, and human approval before actions occur, and they should log blocked or rejected actions as carefully as successful ones. Because artifacts may contain sensitive intelligence such as internal hostnames, headers, or API-related details, they must be sanitized, protected, and eventually archived or deleted. Overall, the chapter frames passive or low-impact reconnaissance as a disciplined first step toward more advanced agentic security systems, preparing the reader for deeper enumeration, scoring, and active testing while keeping accountability and ethics central.

Our passive reconnaissance pipeline. The pipeline gathers information about a target domain and its subdomains. For each discovered subdomain, the pipeline queries DNS servers to retrieve the IP addresses to which the subdomain points. The pipeline then sends an HTTP HEAD request to each live subdomain. If the subdomain uses HTTPS, the pipeline examines the TLS/SSL certificate, checks for Application Layer Protocol Negotiation (ALPN), and inspects the Subject Alternative Name (SAN). All this information gets written to JSONL.

The architecture of a minimal AI recon pipeline. We’ll create a passive recon agent that performs safe actions, collects and interprets the results from these safe actions, and records them to an artifact file.

Each artifact records one cycle of the recon loop. At the Plan stage, the AI agent reads in the seed list and generates candidate subdomains. During the Action stage, it issues DNS queries and HTTPS HEAD requests to each candidate. The Observation stage collects the evidence, which includes resolved Ips, response headers, TLS certificate fields, and WAF hints. Finally, the Record stage persists everything into a JSONL file. Together, these four stages form a complete account of discovery in recon.

• Non-intrusive reconnaissance gathers intelligence about a target through standard, low-impact interactions (DNS lookups, HEAD requests, TLS handshakes) that any browser or crawler might perform. Truly passive techniques such as certificate transparency, Shodan, WHOIS, and archived DNS rely entirely on third-party data and never touch the target at all.
• A pipeline differs from a script in that it connects multiple stages, with each stage's output feeding the next, producing structured artifacts with timestamps and context rather than ad hoc notes.
• The ReAct pattern (Plan, Act, Observe, Record) organizes a recon agent into four reusable stages: generating candidate hostnames, gathering data through DNS and HTTP probes, interpreting the results, and persisting findings as structured artifacts.
• JSONL artifacts serve as persistent, machine-readable memory for the agent, capturing what was done and what was found, forming an audit trail that AI agents can later query, filter, and reason over.
• DNS resolution, HTTPS HEAD requests, WAF/CDN header heuristics, TLS certificate inspection (ALPN and SAN), and robots.txt/sitemap retrieval are the core non-intrusive techniques the pipeline uses to map a target's attack surface.
• Response headers should be sanitized before writing to disk to prevent sensitive data such as session cookies, authorization tokens, and API keys from persisting in log files.
• Safety gates, including time-window checks and approval confirmations, enforce operational boundaries and keep humans in the loop before the agent takes network actions.
• Scope enforcement must be a configured system constraint, not an assumption. Every out-of-scope request should be blocked and logged as an artifact.
• Structured artifacts enable emergent inferences that were never explicitly programmed, because AI agents excel at interpreting consistent, well-formatted data with context rather than unstructured logs.
• The human operator remains accountable for every action the AI system performs; AI suggestions must be explainable, operate under the principle of least privilege, and be fully auditable, whether accepted or rejected.

FAQ