Red Canary Builds Agentic AI Pipeline to Triage Phishing
Red Canary says it built an agentic AI pipeline, a graph of narrowly scoped subagents plus a deterministic rules engine, to triage reported phishing emails, reaching 94% accuracy according to a June 2026 engineering blog post. The pipeline separates parsing and enrichment, hybrid rule-based and LLM-derived feature extraction, a rules engine that can override model outputs, and a final classification step with analyst feedback loops for retraining. Red Canary's product page separately advertises 99% noise reduction and a 60% reduction in triage and notification time across its broader AI agent suite, figures distinct from the phishing-specific 94% number. The company cites Anti-Phishing Working Group data showing more than 1.1 million phishing emails reported in Q2 2025 alone. For SecOps and ML practitioners, the architecture illustrates a reusable pattern: decompose detection into small, auditable components rather than one large model.
For SecOps and ML engineering teams, the most transferable lesson here is architectural, not the specific accuracy figure: Red Canary decomposed phishing triage into narrowly scoped subagents joined by a deterministic rules engine, trading a single large-model approach for modular components that are individually testable, auditable, and can be overridden by hard rules when the model disagrees. That pattern generalizes well beyond phishing to other high-volume, adversarial detection problems.
What happened
Red Canary's engineering blog (published June 2026, author Ryan Morton) describes an AI triage agent built from orchestrated subagents arranged in a graph workflow, reporting 94% accuracy for the end-to-end pipeline. The company's product page separately advertises 99% noise reduction, 3-minute AI agent investigations, and a 60% reduction in triage and notification time; those figures describe Red Canary's broader AI agent suite (which has completed over 2.5 million investigations since a June 2025 launch) rather than the phishing agent specifically. The blog cites the Anti-Phishing Working Group's count of over 3.8 million phishing attacks in 2025, with Q2 alone accounting for more than 1.1 million, the highest quarterly total in two years.
Technical context
The pipeline has four stages, per the blog. A parsing and enrichment subagent normalizes the raw email and adds domain reputation and historical-indicator context. A feature extraction subagent combines traditional boolean checks with AI-powered prompted checks that return true/false judgments plus reasoning, distilling NLP signals like sentiment and intent into simple features. A rules engine applies deterministic overrides before classification, letting Red Canary hand-write rules for emerging campaigns without retraining a model. A classification subagent, trained only on those distilled features (not raw email content or customer data), makes the final call and generates a human-readable explanation.
Industry context
The architecture mirrors a broader shift in SecOps tooling: rather than one large model handling an entire workflow, vendors are increasingly chaining small, specialized agents with deterministic checkpoints. This improves explainability, since rule hits and boolean features produce traceable decision points, at the cost of added orchestration and maintenance complexity - prompt engineering, drift monitoring, and per-environment rule tuning become ongoing engineering work rather than one-time model training.
For practitioners
Teams building similar pipelines should note Red Canary's explicit design choices: classification models are trained only on distilled true/false features, never raw email or customer content, which simplifies data governance; rules can be added ad hoc for emerging campaigns without waiting on retraining; and analyst feedback (thumbs up/down) feeds a continuous tuning loop rather than a periodic retrain cycle.
What to watch
The reported 94% accuracy and the separate 99%/60% product-page metrics have not been independently benchmarked against other SOC vendors' phishing-triage tools. Watch for published false-positive and false-negative rates, latency under production load, and whether Red Canary or competitors release comparative evaluations. Also watch how the rules engine and model outputs are reconciled when they disagree, and whether data-retention practices for enriched email metadata get documented in more detail.
Key Points
- 1Red Canary's agentic phishing-triage pipeline combines deterministic rules with LLM-derived features to reach a reported 94 percent end-to-end accuracy.
- 2The architecture trains classifiers only on distilled true/false features, never raw email content, which simplifies data governance for sensitive traffic.
- 3Practitioners building similar detection pipelines can reuse the pattern: modular, testable subagents plus rule-based overrides for cases models get wrong.
Scoring Rationale
A concrete, now fully source-verified practitioner case study of a hybrid deterministic-LLM agentic pipeline for phishing triage, with real architectural detail. Genuine value for SecOps and ML engineers, but the core 94% accuracy figure comes only from Red Canary's own blog with no independent benchmark, and the 99%/60% headline metrics describe the company's broader AI suite, not the phishing agent specifically.
Sources
Public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problems
