AI-Assisted Phishing: Flawless Lures at Infinite Scale

AI-assisted phishing uses large language models to write grammatically clean, context-aware, localised phishing and business-email-compromise (BEC) lures at scale. It removes the spelling, grammar and translation errors that traditionally flagged phishing, and enables rapid persona and pretext generation, target-reconnaissance summarisation, and iterative refinement of lures. The effect is more volume and more polish — and a lower skill barrier for novice actors — rather than a fundamentally new technique. The UK NCSC assessed in January 2024 that AI will “almost certainly” increase the volume and impact of attacks, with the biggest uplift in social engineering.

How It Works

Attackers prompt mainstream or jailbroken LLMs to draft lures, translate them into many languages, generate convincing sender personas, summarise reconnaissance on a target, and A/B-refine messaging. AI is an enabling tooling layer on top of existing techniques: spearphishing by link, attachment or voice, impersonation, and account creation. T1566 T1598 T1656 T1585

What the Data Shows

Measured rates are hard to pin down because AI-text detectors are unreliable, so published figures are vendor estimates, not ground truth. Security vendors using such detectors have estimated that a meaningful and rising share of BEC and spam email is AI-generated (for example, Barracuda’s analysis through April 2025 estimated roughly half of spam and a growing minority of BEC). Treat specific percentages as directional. The consistent, higher-confidence signal is the NCSC’s: a large uplift in social-engineering volume and quality, and a significant boost for less-skilled actors.

Detection & Mitigation

• Shift detection away from grammar and spelling heuristics toward authentication and behavioural signals: DMARC/DKIM/SPF, display-name and look-alike-domain analysis, and anomalous payment-change requests.
• Deploy phishing-resistant MFA (FIDO2/passkeys) so credential-harvest lures do not convert to account takeover.
• Require out-of-band verification for any payment or banking-detail change (the core BEC control).
• Update awareness training to assume flawless, personalised, multi-language lures; use AI-assisted filtering and URL/attachment detonation, and monitor for mass-personalised send patterns.