Email security is defined by impermanence: a perpetual call-and-response arms race in which defenses are only as good as the last bypass discovered. Every control eventually becomes yesterday's solution. The harder problem is that the biggest weaknesses are, by definition, invisible.
Abraham Wald's World War II work on bomber armor illustrates the issue. Engineers wanted to reinforce the areas showing bullet holes on planes that returned; Wald pointed out they were reinforcing the places aircraft could already survive damage. The real vulnerabilities sat on the planes that never came back.
Phishing detection has the same blind spot: the gaps in coverage are unseen. Cloudflare has integrated LLMs into its email security tooling to move from reactive remediation to proactive improvement.
Why user reports are a lagging indicator
Traditional systems improve mainly through user-reported misses. When a spam message is classified as clean, customers can send the original EML into the pipeline for analysts to review and models to be updated. That feedback loop is necessary and valuable, but it is inherently reactive — it depends on someone noticing a failure after the fact and taking the time to report it. Detection improvements therefore tend to be driven by what attackers have already made work, not by what they are about to exploit.
Mapping the threat landscape with LLMs
Large language models, mainstream since late 2022 and early 2023, use deep learning and massive datasets to predict the next token in a sequence, which lets them understand context and nuance. That makes them well suited to email: they can read natural language and characterize complex concepts such as intent, urgency and deception across millions of messages.
Cloudflare processes millions of unwanted emails daily. Historically it was not feasible to characterize each message beyond coarse classifications, and manually mapping emails to nuanced threat vectors did not scale. LLMs now generate tags that give analysts high-fidelity signals in near real time, with relevant context attached. Work that previously took hours of manual investigation and complex querying is surfaced automatically, increasing the speed at which new targeted ML models can be built or existing ones retrained. Because Cloudflare operates at global Internet scale, these insights often arrive before a new technique shows up widely in customer-reported misses.

LLM-driven categorization shows clear spikes and persistent trends across categories including "PrizeNotification" and "SalesOutreach".
Case study: Sales Outreach phishing
One of the clearest patterns surfaced by the new intelligence is the persistence of malicious messages built to resemble Sales Outreach. These mimic legitimate B2B communication — often offering to purchase or receive "special deals" on unique items or services — to lure targets into clicking malicious links or handing over credentials.
Once categorization flagged Sales Outreach as a dominant vector, Cloudflare moved from broad visibility to targeted data collection, isolating messages with those characteristics across its global dataset. The result was a continuously growing, high-precision corpus of real-world examples: confirmed malicious messages alongside borderline cases traditional systems struggled to classify. A dedicated training pipeline was built from it:
- Data curation: messages grouped by shared linguistic and structural traits identified by the LLMs — persuasive framing, manufactured urgency, transactional language and subtle social proof.
- Feature extraction: focused on sentiment and intent rather than static indicators, so the model learns how requests are phrased, how credibility is established and how calls to action are embedded in otherwise normal business conversation.
- Model training: a purpose-built sentiment analysis model optimized for Sales Outreach behavior, avoiding overloading a general phishing classifier and allowing precision and recall to be tuned for this threat class.

From language to enforcement
The model outputs a risk score reflecting how closely a message matches known Sales Outreach attack patterns. That score is evaluated alongside sender reputation, link behavior and historical context to decide whether a message is blocked, quarantined or allowed. The process is continuous: as attackers adapt their language, newly observed messages feed back into the pipeline to refine the model without waiting for large volumes of user-reported misses. LLMs act as the discovery layer, surfacing new linguistic variants, while the specialized model handles fast, scalable enforcement.
Measured impact
The visibility unlocked by LLM-driven mapping changed how detections are improved: instead of waiting for attackers to succeed and relying on downstream reports, systemic gaps can be identified earlier and fixed at the source.
The clearest signal was reduced customer friction. Sales Outreach phishing has historically generated a high volume of reported misses because it closely resembles legitimate business communication and often evades rule-based or reputation-driven systems. Average daily Sales Outreach submissions — messages labeled clean that were in fact Sales Outreach phishing, flagged by end users — fell from 965 in Q3 2025 to 769 in Q4 2025, a 20.4% reduction in reported misses in a single quarter. The trend continued in Q1 2026, with average daily submissions down by two-thirds.

Each avoided submission is a phishing attempt stopped before it could erode trust, consume analyst time or force a user to make a security judgment mid-workflow.
What comes next
Cloudflare describes its LLM work as just beginning. The next step is total environmental awareness: refining LLM specificity to extract forensic-level detail from every interaction, so specific tactical signatures can be identified rather than relying on broad labels. In parallel, specialized machine learning models are being deployed to hunt emerging, highly obfuscated vectors at the fringes where traditional defenses miss. Real-time LLM data serves as a strategic compass, shifting human expertise away from known noise and toward the gaps where the next strike is likely to land.
Organizations can access Cloudflare's Retro Scan tool free of charge, whether or not they are customers. It uses the company's predictive AI models to scan existing inbox messages in Microsoft 365, highlight threats and enable remediation directly in the email accounts, after which organizations can apply further controls to prevent similar messages from arriving.




