AI Just Got Very Good at Lying to You — And Your Business Is the Target | Robert Braxton

A WIRED experiment showed five AI models running convincing social engineering attacks — completely autonomously. The 'human risk' that causes 90% of enterprise breaches just got automated. Here's what Australian businesses need to do right now.

WIRED journalist Will Knight recently watched an AI model craft a personalised phishing email targeting him — referencing his newsletter, his published work, his specific areas of interest — then autonomously respond to his replies in real time to keep him engaged and inch him toward clicking a malicious link. The model doing this was DeepSeek-V3. The tool running it was built by a startup called Charlemagne Labs. The experiment tested five different AI models including Claude 3 Haiku, GPT-4o, Nvidia Nemotron, DeepSeek V3, and Alibaba's Qwen. All of them attempted convincing social engineering attacks. Some were better than others. None of them required a human operator once the attack was initiated. This is the moment I've been preparing clients for at Cyborium for the past two years. It has arrived. What "human risk" actually means in 2026** Jeremy Philip Galen, co-founder of Charlemagne Labs, put the core problem clearly: "The genesis of 90 percent of contemporary enterprise attacks is human risk." That number hasn't changed much in 20 years. What has changed is the cost and scale of exploiting it. Until recently, a convincing, personalised phishing campaign required skilled human operators — researchers who could find personal details, writers who could craft believable messages, operators who could manage real-time back-and-forth with targets. That work was expensive and time-consuming, which meant attackers had to pick targets selectively. AI has eliminated that bottleneck entirely. What previously required a team and weeks of preparation can now be done by a single operator running automated pipelines across thousands of targets simultaneously. Rachel Tobac, CEO of SocialProof Security, described it directly: "I wouldn't say AI has made attacks more convincing, but it has made it easier for one person to scale attacks. The kill chain is getting entirely automated." Automated. The entire kill chain. Let that register. Why AI is particularly effective at this** The same quality that makes AI chatbots frustrating in other contexts — sycophancy, the tendency to flatter and agree with the person they're speaking to — makes them dangerously effective social engineers. A good social engineering attack isn't built on pressure; it's built on rapport. It mirrors the target's interests, validates their judgment, and creates a sense of shared purpose before making the ask. AI models do this naturally. They're trained to be agreeable and engaging. An attacker pointing that capability at a human target, with specific research about that person loaded as context, produces emails and messages that feel personal, relevant, and trustworthy. The WIRED experiment showed AI models doing this cold, with minimal setup. With a motivated attacker doing even basic LinkedIn research first, the results would be significantly more convincing. What this means for your business** If your team uses email, Slack, Teams, LinkedIn messages, or any other digital communication channel — and every business does — this threat applies to you directly. Here's how I'd frame the risk profile for Australian SMEs: High-value targets within your business**: Finance team members who authorise payments. Anyone with admin access to cloud systems. HR staff who handle personal data. Executives who communicate with investors, lawyers, or board members. These people are research targets. Their email addresses, roles, and communication patterns are often visible through LinkedIn, your website, and public records. Common attack vectors**: Fake vendor onboarding emails. Fraudulent invoice approval requests (business email compromise). Fake IT support messages requesting credential resets. LinkedIn connection requests from synthetic profiles preceding a targeted message campaign. The scale problem**: Previously, your team might receive a handful of targeted phishing attempts per year. With automated AI pipelines, that could become hundreds per week, each personalised to specific team members based on their public digital footprint. Five things to do before next Monday** 1. **Run a social engineering awareness session with your team.** Show them the WIRED article. Make the threat concrete. Most people don't understand that a convincing, personalised email can now be generated by an AI in seconds. Awareness is the first line of defence. 2. **Implement a verification protocol for financial requests.** Any email requesting a payment, account change, or credential update should require a secondary verification through a known phone number — not a reply to the same email thread. This one control stops the majority of business email compromise attacks. 3. **Audit your public digital footprint.** Search your team members on LinkedIn and your own website. The more detail that's publicly visible about roles, responsibilities, and reporting lines, the more useful it is to an attacker building a targeting profile. Consider what needs to be public and what doesn't. 4. **Enable AI-assisted email filtering.** Modern email security platforms now include AI-powered phishing detection that can identify AI-generated messages. If you're still relying on basic spam filters, you are not adequately protected against this threat class. 5. **Brief your finance and IT teams specifically.** These are the two highest-value targets in most SMEs. Finance because of payment authorisation. IT because of system access. Both groups should have explicit protocols for verifying unusual requests, regardless of how legitimate the sender appears. On the open-source defence question** The WIRED article closes with an important point from Richard Whaling, who co-founded Charlemagne Labs: "Having powerful models on the defensive side of the fence may outweigh the risks. We rely on open-source models to train our defensive model. That relies on a healthy open-source community. And that might be the only viable way to defend ourselves." He's right. The same AI capabilities being weaponised by attackers are available to defenders — for email filtering, anomaly detection, user behaviour analysis, and automated threat response. The organisations that invest in defensive AI capabilities now will be materially better positioned than those that don't as this threat scales through 2026 and beyond. At Cyborium, we're building these capabilities into the security programs we deliver for Australian businesses. The threat is real. The defences exist. The gap between organisations that have implemented them and those that haven't is widening every month. Don't be on the wrong side of that gap. Robert Braxton is CEO of Simple Complex Solutions and Executive Director at Cyborium, an Australian cybersecurity firm specialising in SME security programs. For a security assessment of your business, visit robertbraxton.com or call +61 417 226 549.*