When the people building the future of artificial intelligence start changing their daily security habits, it’s worth paying attention. Recent security advice for employees at OpenAI and Anthropic has highlighted a new level of real-world risk.
A New Kind of Threat
The guidance
gaining attention isn't just about complex cybersecurity measures; it's about physical and personal security. Employees at both OpenAI, the creator of ChatGPT, and Anthropic, the company behind Claude, have reportedly been advised to conceal company-branded clothing and hide their work badges when in public. This isn't just corporate caution. It stems from a growing number of incidents, including protests, targeted hostility, and even individuals attempting to spy on employees to gain insider information. One report mentioned a spy approaching an OpenAI employee at a playground, trying to steer the conversation toward their work. These events underscore a significant shift: working for a top AI company now makes you a potential target, not just for digital attacks, but for in-person espionage and harassment.
Espionage in the AI Era
The threat of corporate and state-sponsored spying is a major driver of the new security posture. OpenAI's security training explicitly warns staff that spies might use tactics ranging from bribery to feigned romantic interest to extract confidential information. The concern is that the race for AI dominance has made employees valuable sources of proprietary knowledge. An investor in both companies noted that people constantly approach AI workers hoping to get firsthand information on cutting-edge developments. This has led to internal policies that go beyond standard digital hygiene, reminding staff to use privacy screens on public transport and to always erase whiteboards in meeting rooms. The guidance reflects a tense reality where a casual conversation in a cafe could be an intelligence-gathering operation.
Internal Security Under Scrutiny
Beyond external threats, the internal security practices of these AI labs have also drawn attention. For years, some OpenAI employees reportedly raised concerns about security weaknesses, worrying that the company's 'move fast' culture was prioritised over protecting its technology from insider threats or foreign governments. These concerns were amplified by incidents where AI models themselves were used to breach other systems, demonstrating how the very tools they are building can be weaponised. In response, companies are beefing up their internal protocols. Both Anthropic and OpenAI are implementing stricter access controls, better software supply chain security, and more robust monitoring to prevent unauthorised access and data leaks. This includes sandboxing, or running AI tests in isolated environments, to prevent unintended consequences.
Advice for the Rest of Us
While most of us don't have to worry about spies seducing us for trade secrets, the principles behind this heightened security are relevant to everyone. The core message from security experts in the AI field is to practice strong foundational hygiene. This includes using approved and vetted AI tools for work, being extremely cautious about what data you input, and establishing clear policies around AI use. Employees are advised to never enter confidential, financial, or sensitive personal information into public AI platforms unless explicitly approved. The same AI that can write a poem can also process and store the sensitive data you provide it. Thinking before you prompt is becoming as critical as thinking before you click.
The Rise of AI in Defence
Ironically, the best defence against AI-powered threats may be AI itself. OpenAI has been vocal about using its own models to create a 'defence factory' to find and fix vulnerabilities in its systems before attackers can exploit them. The company recently announced GPT-6 Cyber, a model specifically trained to find and validate security issues. This approach, known as red teaming, involves using AI agents to simulate sophisticated attacks to identify weaknesses. The goal is to create a system of continuous defence where AI helps secure the very software it is built upon. This reflects a broader industry belief that defenders have a limited window to use AI to get ahead of attackers.
