The inbox has long been the softest entry point in enterprise security. As phishing campaigns grow more convincing, more personalised, and increasingly powered by generative AI, the tools designed to stop them have been locked in a reactive cycle: wait for the attack, analyse it, respond. IRONSCALES, the Atlanta-based email security vendor, is betting that cycle is about to break.
Ahead of this week’s RSA Conference in San Francisco, the company announced a new threat intelligence initiative alongside live demonstrations of the three AI agents it shipped in its Winter 2026 platform release. Together, the moves represent IRONSCALES’ push to reposition itself from a detection vendor into something closer to a preemptive security partner, one that models attacks before they arrive rather than cataloguing them after the fact.
What the new intelligence series actually does
The “Email Attack of the Day” series, which IRONSCALES is debuting at RSAC 2026, draws on anonymised threat data from its network of more than 17,000 customer organisations. The concept is straightforward: surface real-world email attack patterns as they emerge, publish them with technical context, and give security teams the intelligence to recognise new tactics before they proliferate.
It is not an entirely novel format. Other vendors publish threat advisories and campaign breakdowns routinely. But IRONSCALES is framing the series as a complement to its broader shift toward what it calls “Phishing 3.0” defences, where intelligence feeds directly into adaptive detection rather than sitting in a separate research silo.
Three AI agents, one architecture
The 💜 of EU tech
The latest rumblings from the EU tech scene, a story from our wise ol' founder Boris, and some questionable AI art. It's free, every week, in your inbox. Sign up now!
The centrepiece of the RSAC demonstrations will be the three AI agents IRONSCALES introduced in its Winter 2026 release earlier this month: Red Teaming, Phishing SOC, and Phishing Simulation. Each is purpose-built rather than layered on top of a general-purpose large language model, a design choice Audian Paxson, principal technical strategist at the company, has described as more efficient for encoding domain-specific expertise.
The Red Teaming agent performs continuous reconnaissance against an organisation’s public footprint, scanning everything from social media presence to executive communications and org charts. It then generates tailored attack simulations and feeds them into the platform’s detection models. The idea is to harden defences against the specific phishing campaigns an attacker would build for that particular organisation, not just the generic threats circulating broadly.
The Phishing SOC agent, meanwhile, handles forensic investigation of suspicious emails. IRONSCALES says it delivers what amounts to a Level 2 analyst’s assessment in minutes, examining five investigative tracks and producing a verdict that would otherwise consume hours of human analyst time. For managed service providers juggling dozens of client environments, the speed difference matters.
The third agent, Phishing Simulation, takes the reconnaissance data gathered by its Red Teaming counterpart and uses it to create hyper-personalised training simulations. Rather than recycling generic phishing templates, it targets an organisation’s highest-risk employees with scenarios drawn from real OSINT data and delivered in their native language.
The wider context: an arms race that favours the attacker
IRONSCALES is making these moves against a backdrop that has grown considerably more hostile. According to research cited in the company’s own announcements, 88 per cent of organisations report falling victim to AI-powered security incidents within the past 12 months. KnowBe4’s 2025 Phishing Threat Trends Report found that more than 82 per cent of phishing emails analysed contained indicators of AI assistance. A Hoxhunt analysis documented a 14-fold surge in AI-generated phishing over the 2025 holiday period alone.
The economics have shifted, too. Where crafting a convincing spear-phishing campaign once required time and skill, generative AI has compressed the effort to minutes and a handful of prompts. IBM security researchers demonstrated that AI could build a phishing campaign as effective as one created by human experts, needing just five prompts instead of 16 hours of work.
RSAC 2026 itself reflects this anxiety. Agentic AI, the category of autonomous systems capable of planning and executing multi-step operations, dominates this year’s conference agenda. Microsoft’s keynote addresses securing AI agents at enterprise scale. Multiple vendors are unveiling deepfake detection tools. The conversation has moved decisively from whether AI will reshape email security to how quickly defenders can close the gap.
Encryption and deepfake protection round out the release
Beyond the AI agents, the Winter 2026 release includes integrated email encryption for outbound messages, a feature IRONSCALES designed to address compliance requirements without adding friction. The system applies encryption through two modes: policy-based protection for regulated content and user-initiated encryption for sensitive workflows.
The release also extends the company’s deepfake protection for Microsoft Teams, which IRONSCALES first introduced in 2025. Enhanced voice detection now learns employee voice patterns passively from normal meeting participation, flagging impersonation attempts even when cameras are switched off. It is a notable addition given that deepfake-driven fraud increased more than 700 per cent year over year, according to Cyble’s 2025 Executive Threat Monitoring data, and Gartner surveys indicate that 62 per cent of organisations experienced a deepfake attempt in the past year.
From reactive to preemptive, at least in theory
The underlying pitch from IRONSCALES is a closed-loop architecture: reconnaissance feeds detection, detection feeds training, and training feeds back into better recognition. Eyal Benishti, the company’s CEO, has described the approach as distinct from competitors who use OSINT-driven attack generation solely for employee training. IRONSCALES, he argues, uses it to improve detection first.
Whether that distinction proves meaningful in practice will depend on how the agents perform at scale across diverse customer environments. The email security market is crowded, and the claim of preemptive protection is one that several vendors are now making simultaneously. But the architectural bet, purpose-built agents feeding a shared adaptive model trained on data from 17,000 organisations, is at least a testable proposition.
Attendees at RSAC 2026 can see the platform demonstrated live at Booth #4600 in the North Expo. For everyone else, the real test will be whether the next wave of AI-powered phishing campaigns encounters defenders who saw them coming.
Get the TNW newsletter
Get the most important tech news in your inbox each week.