The anonymization playbook enterprises have relied on for a decade is functionally broken. Gartner forecasts that by 2029, most privacy incidents will originate from AI-generated inferences rather than direct exposure of names, Social Security numbers, or medical records. A 2026 study found that LLMs can reidentify pseudonymous individuals from online profiles alone, achieving 68% recall at 90% precision where non-LLM methods scored near zero. Cobalt CISO Andrew Obadiaru puts the operational threat plainly: AI can now infer an employee’s organizational influence or phishing susceptibility from publicly available, apparently innocuous data<\/a> without touching a single protected file.<\/p>\n The enterprises most exposed here are not the ones with the worst breach history. They are the ones with the longest data retention tails, the most permissive internal data sharing, and the most public-facing digital presence across LinkedIn, GitHub, and marketing databases. If your organization has spent the last five years investing in perimeter security and GDPR compliance as the primary privacy posture, you have been optimizing against a threat model that AI has already made obsolete. The question is not whether you have a data breach, but what an adversary can infer about your org chart, spending authority, and critical system owners from data you consider non-sensitive.<\/p>\n The underlying shift is one of economics, not just capability. Obadiaru’s observation that building detailed profiles of thousands of targets was not economically viable five years ago but is today is the sharpest line in the piece. Targeted social engineering has historically been expensive and slow, which meant attackers rationed it against the highest-value targets. AI inference eliminates that constraint. Business email compromise and phishing campaigns that used to require hours of manual research per target can now be industrialized. That changes the risk calculus for mid-market enterprises and second-tier vendors in supply chains who previously assumed they were below the effort threshold.<\/p>\n Gartner’s Bart Willemsen recommends privacy-enhancing technologies like differential privacy and homomorphic encryption, which processes calculations on encrypted data without decrypting it first, alongside aggressive data lifecycle enforcement. Those are the right tools, but the harder governance problem is cultural. Most enterprises collect data without a deletion mandate, treat retention as free, and have never mapped what a sufficiently capable AI could infer by correlating their HR directory against their public GitHub activity against their customer interaction logs. The CISO who hasn’t run that thought experiment against their own estate is the one an attacker will run it against first.<\/p>\n An inference attack is when an AI system reconstructs sensitive information that was never explicitly shared, by combining and pattern-matching data points that individually seem harmless. Think of it as triangulation: knowing someone’s job title, their commute time from a check-in app, and three purchase categories from a loyalty program is enough for a model to estimate their income bracket, health conditions, or relationship status. The business risk is that no single data asset needs to be breached for the attack to succeed.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: AI inference attack<\/h2>\n