PhishLumos: An Adaptive Multi-Agent System for Proactive Phishing Campaign Mitigation

Technical Summary: PhishLumos

Problem Statement
Phishing attacks represent a persistent societal threat, causing significant financial loss and eroding trust in digital services, particularly among vulnerable populations. Current defensive strategies rely heavily on reactive, content-based detection methods (e.g., visual similarity, logo recognition, and text analysis). However, modern attackers employ sophisticated evasive tactics, such as cloaking, which deliver malicious content to human users while presenting benign or empty content (e.g., 404 errors) to automated scanners. This renders traditional content-based detectors ineffective. The paper identifies a core cybersecurity asymmetry: attackers can easily scale coordinated campaigns, while defense remains a resource-intensive, expert-driven task. Furthermore, the prevalence of content-inaccessible scenarios has escalated, with 100% of campaigns observed in early 2025 failing to provide retrievable content to automated scanners.

Methodology: PhishLumos
The authors introduce PhishLumos, an adaptive multi-agent system designed to proactively mitigate entire phishing campaigns starting from a single URL, even when the page content is inaccessible. The system operates on the principle of Pivotal Analysis, treating content unavailability not as a failure, but as a signal to pivot the investigation toward the attacker's underlying infrastructure.

• Architecture: The system is orchestrated by a central Supervisor Agent that directs a team of specialized and synthesis agents.
  • Specialized Agents: These agents (URL Analysis, Domain Analysis, IP Analysis, Certificate Analysis) query external threat intelligence APIs (e.g., urlscan.io, VirusTotal) to gather historical DNS records, IP address histories, and TLS certificate data.
  • Synthesis Agents: The Campaign Analysis Agent correlates findings to profile the campaign, while the Rule Generation Agent creates and validates detection rules.
• Adaptive Workflow: The Supervisor uses a Large Language Model (LLM) to evaluate collected evidence and dynamically decide the next investigative step (Algorithm 1). If initial scans reveal cloaking or suspicious infrastructure patterns (e.g., multiple unrelated domains on a non-CDN IP), the system pivots to analyze shared hosting, certificates, or registration patterns.
• Context Management: To handle LLM context limits, the system employs a strategy of selective data extraction, structured inter-agent communication, and hierarchical summarization to distill raw API data into concise evidence graphs.
• Rule Generation: The system generates candidate detection rules (e.g., infrastructure-based queries) and empirically validates them against a small data sample to ensure high precision and prevent false positives before outputting actionable intelligence.

Key Contributions

1. Shift from Reactive to Proactive: PhishLumos moves beyond blocking individual URLs to mitigating entire attack campaigns by identifying shared infrastructure footprints.
2. Resilience to Evasion: Unlike prior work that degrades when content is hidden, PhishLumos leverages infrastructure signals (IPs, certificates, ASNs) to detect campaigns even when page content is cloaked or inaccessible.
3. Multi-Agent Orchestration: The system demonstrates a novel application of LLM-powered agents to coordinate complex, state-aware investigations, dynamically shifting focus based on emerging evidence.
4. Actionable Intelligence: The system outputs validated detection rules (e.g., search queries for blocklists) rather than just binary classifications, enabling direct integration into security ecosystems.

Experimental Results
The system was evaluated on a ground-truth dataset of 103 distinct phishing campaigns (6,020 URLs) verified by JPCERT/CC, alongside a benign dataset.

• Campaign Coverage: PhishLumos achieved a median campaign coverage of 100%, meaning the generated rules successfully identified nearly all known URLs within a campaign.
• Proactive Detection: The system identified campaigns a median of 192.8 hours (approx. 8 days) before human experts confirmed them, providing a critical window for preemptive defense.
• Performance in Evasive Scenarios: In content-inaccessible scenarios (85 campaigns), PhishLumos achieved an F1-Score of 0.994 with 1.000 recall. In contrast, state-of-the-art baselines (ChatPhishDetector, Phishpedia, VisualPhishNet) saw their performance collapse, with F1-scores dropping to as low as 0.000 for visual-based methods.
• Precision: The system maintained a minimal False Positive Rate of 0.001 across all scenarios.
• Ablation Study: Removing the "Pivotal Analysis" capability reduced mean coverage from 93.0% to 56.1%, confirming the necessity of infrastructure pivoting. Removing the "Rule Generation" agent (specifically the multi-hypothesis validation step) caused median coverage to plummet to 2.2%, highlighting the importance of rigorous rule validation.

Significance
The paper claims that PhishLumos represents a practical shift in cybersecurity defense. By automating the expert-level investigation of campaign infrastructure, it addresses the scalability gap between attackers and defenders. The system's ability to detect campaigns days before official verification allows security vendors and organizations to deploy blocking rules network-wide, protecting vulnerable users before they are harmed. The work demonstrates that effective defense against modern, evasive phishing does not require access to malicious content, but rather the ability to reason about the persistent infrastructure that supports these campaigns.