AI-Powered Phishing: How Threat Actors Are Using LLMs to Scale Spearphishing
Large language models have eliminated the grammar mistakes and awkward phrasing that made phishing emails easy to spot. Here is what the new generation of AI-generated phishing looks like and how to defend against it.
Introduction
For years, security awareness training taught employees to spot phishing by looking for poor grammar, awkward phrasing, and generic greetings. That heuristic is now largely dead. Large language models have given threat actors the ability to produce native-quality written content in any language, at scale, personalised to the target.
What Has Changed
The traditional spearphishing workflow required a skilled operator to manually research a target, craft a convincing pretext, and write a believable email. This took time and limited the scale of operations. LLMs collapse that bottleneck entirely.
An attacker can now feed a model a LinkedIn profile, a company website, and recent news about the target organisation and receive a highly contextualised, grammatically perfect phishing email in seconds. The same workflow can be automated to target thousands of individuals simultaneously.
What AI-Generated Phishing Looks Like
The tells have changed. Look for these patterns instead of grammar mistakes:
Unusual precision. AI-generated emails often reference very specific details about the target — recent projects, job title changes, mutual connections — that feel uncanny coming from an unsolicited email.
Perfect but sterile tone. The writing is flawless but lacks the personality of someone who actually knows you. There are no in-jokes, no shared references, no imperfections.
Urgency without context. LLMs are good at creating urgency but struggle to make it feel organic. Watch for time pressure that does not connect naturally to the stated scenario.
Real Campaigns Observed
In 2024, multiple threat intelligence vendors documented campaigns using LLM-generated content targeting financial services executives. The emails referenced real earnings calls, used correct internal terminology, and impersonated known counterparties with high fidelity.
One campaign targeting a European bank used LLM-generated voice synthesis alongside email — the victim received a phishing email followed by a phone call from what sounded like a known colleague requesting urgent wire transfer approval.
Defensive Measures
Assume content quality is no longer a signal. Retrain your workforce. A well-written email is not a safe email.
Focus on metadata and behaviour. Email headers, sending infrastructure, domain age, and link destinations are harder to fake than content.
Implement DMARC, DKIM, and SPF strictly. LLMs can write perfect emails but they cannot authenticate as your CEO's actual mail server.
Out-of-band verification for high-risk actions. Any request involving wire transfers, credential resets, or sensitive data access should require a separate verified communication channel regardless of how convincing the initial request appears.
Conclusion
AI has not invented new attack categories. It has removed the skill floor from existing ones. Phishing that previously required a trained social engineer can now be produced by anyone with API access. Defences need to shift from content analysis to infrastructure analysis and process controls.