AI’s expanding attack surface

How the rush to deploy AI is creating new cybersecurity risks

A recently disclosed security flaw, dubbed “Echoleak”, showed how an AI assistant embedded in office software could be induced to exfiltrate sensitive data without any user interaction. The vulnerability was swiftly addressed. But as AI applications multiply across enterprises, and as attackers become more creative, such episodes are unlikely to remain isolated.

Without clear rules about who can access which data, and for what purpose, AI can quickly get into trouble. It may follow instructions from the wrong person, overshare data or expose firms to unfamiliar forms of risk. Those dangers are amplified by the sheer number of systems now in use.

“AI is so transformational that businesses are going to do this, whether it’s secure or not,” says Adam Geller, chief product officer at cloud security firm Zscaler. Many chief information security officers (CISOs), he says, take a pragmatic view: they aim to “de-risk” AI adoption but are “not standing in the way of that train”.

"We're in an AI gold rush. Business leaders are feeling the pressure to innovate."

Jennifer Gold, CISO at Risk Aperture

Jennifer Gold, CISO at Risk Aperture, a cybersecurity company founded by former intelligence community leaders, sees a similar acceleration. “We’re in an AI gold rush,” she says. “Business leaders are feeling the pressure to innovate.”

At the same time, more organisations are deploying public-facing AI systems that may have security vulnerabilities, potentially putting sensitive corporate and customer data at risk. Each new AI interface or integration expands the attack surface that cybercriminals can exploit and defenders must secure.

A wider scope for exploitation

Attackers have been quick to take advantage of the shift. In August 2025, Anthropic said that its Claude AI assistant had been misused in extortion schemes, fraudulent employment operations and the sale of AI-generated ransomware.1 Two months later, it reported what it described as the first documented AI-orchestrated cyber-espionage campaign, in which threat actors used agentic AI to execute 80-90% of tactical operations with minimal human oversight.2

Organisational inertia often compounds the problem. The rapid adoption of AI agents, which perform actions and make decisions autonomously, adds a further layer of complexity.

Young Woman Looking at Big Digital Screens Glitching While Displaying Code Lines

Ms Gold advises treating AI assistants like new teammates with clearly defined roles. AI tools should only have access to information or connections to other tools with the appropriate credentials. “We would never grant a new, untrained employee unfettered access or decision-making authority,” she says. A new AI agent—just like a new employee—might be helpful and efficient, Ms Gold explains, “but have they been trained to fully understand all of these adversarial patterns and how they can be compromised?”

Ensuring the safe development and deployment of AI applications increasingly requires testing throughout the system’s life cycle, rather than by checks applied only at launch. In response, some security firms are expanding their capabilities to focus specifically on how AI models, agents and workflows behave under attack.

Zscaler recently acquired SPLX, a company specialising in stress-testing AI systems through techniques such as red teaming (which involves the simulation of real-world, sophisticated attacks) and prompt inspection. The aim is to give organisations greater visibility into where AI is being used, how it interacts with data and other tools, and where vulnerabilities may emerge as those systems evolve.

Technical controls alone, however, are only part of the challenge.

Team looking at computerYoung woman with data exposure

Shadow AI and human risk

Employee misuse—deliberate or unintentional—of public AI is another new risk that companies must address. Employees may paste sensitive material into public tools or use unsanctioned apps, leading to accidental data leaks. In 2023, Samsung suspended access to generative AI after discovering that staffers had uploaded proprietary code to ChatGPT.3 The incident was an early reminder of how quickly AI can turn small cracks into major incidents.

Mr Geller cautions that simply blocking unapproved AI apps can create a false sense of security. Usage may appear to fall, but in reality it often shifts out of sight. A more effective approach, he says, is to provide a secure, seamless, single-sign-on alternative that matches employee expectations for convenience, usability and speed.

At Zscaler, this thinking underpins tools such as AI Guard, which is designed to function as a checkpoint between a user and AI. By monitoring prompts, data uploads and responses from AI apps in real time, and by blocking unsafe interactions, the tool aims to reduce the risk of sensitive information being exposed or harmful outputs being generated.

"Security and productivity don’t have to be at odds."

Adam Geller, chief product officer at cloud security firm Zscaler

The architecture of trust

There is no need to halt the use of AI, says Mr Geller, but there is a need to shape its adoption in a responsible way. “Security and productivity don’t have to be at odds,” he points out.

AI can act as a risk multiplier. With clear guardrails, however, it can strengthen defences. “At the end of the day, we are meant to be enabling businesses,” Mr Geller says of himself and his cybersecurity colleagues. “The business accelerates with AI; security has to accelerate faster or else become the failure point.”

References

 1. Anthropic, “Detecting and countering misuse of AI,” August 27th 2025.

2. Anthropic, “Disrupting the first reported AI-orchestrated cyber espionage campaign,” November 17th 2025.

3. Bloomberg, “Samsung Bans Staff’s AI Use After Spotting ChatGPT Data Leak,” May 2nd 2023.