The New Cyber Battlefield: AI vs AI and the Rise of Autonomous Security Systems
The Security StrategistShow Notes
The moment an organisation's board starts asking how to prepare for autonomous AI attacks, the conversation has already shifted. What used to be a theoretical briefing topic is now a line item in risk registers and a direct question landing on CISOs' desks from the C-suite. Shachar Hirshberg and Dan Shiebler, co-founders of Artemis Security, an AI-Native Protection Platform for security operations, in production at Mercury, Lemonade, Wix, Upwork, and some of the largest enterprises in the world, have that conversation daily. Artemis raised $70M in series A, led by Felicis with First Round Capital and Brightmind Partners doubling down, alongside top VCs including Theory Ventures, Lockstep, Two Sigma Ventures, and prominent cybersecurity industry leaders, including the founders of Abnormal AI and Demisto, the former CEO and CTO of Splunk, and senior executives from CrowdStrike, Palo Alto Networks, Microsoft, and Okta. In a recent episode of the Security Strategist Podcast with host Richard Stiennon, Hirshberg and Shiebler laid out the strategic reality with unusual clarity, not as a product pitch, but as a candid assessment of where the threat environment stands and what it demands from security leadership. The Economics of Attack Have Changed The foundation of legacy security architecture rests on an assumption that no longer holds: that launching a sophisticated, targeted attack is expensive. Acquiring intelligence on a specific organisation, crafting adaptive exploits, and manually steering a multi-stage breach required time, skill, and resources. Defenders could lean on that cost. Understand attacker behaviour, get ahead of their patterns, and you impose meaningful friction. Shiebler identifies this as the core structural failure of traditional approaches today. "AI really changes that. It's so much easier for attackers to craft new attacks, to explore different strategies, and make it much cheaper to send out radically different, really sophisticated attacks, which really means that trying to rely on approaches that involve just understanding attackers and trying to stay ahead of that is very, very challenging." The consequence is not simply faster attacks. It's the collapse of the distinction between opportunistic, broad-based threats and sophisticated targeted campaigns. What previously required nation-state resources or advanced persistent threat infrastructure can now be approximated by an attacker with limited technical knowledge and access to capable agentic tooling. The MTTR Calculation Hirshberg frames the urgency in operational terms. The industry benchmark for mean time to respond sits at roughly four hours. The top 0.1 per cent of security operations globally measure in minutes. The frontier measures in seconds and adversaries are already in seconds. "We are still talking in hours and need to bridge that gap because we will live in an era where it will have a hundred real zero days every single day in every organisation. If you're measuring your MTTR in hours and you have a hundred real attacks per day, you are fully overwhelmed with traditional tooling." The arithmetic is unambiguous, and no staffing model resolves it. No incremental tooling investment closes it. It requires a categorical shift in how detection, investigation, and response are architected, moving from human-executed to human-guided autonomous response. The Defender’s Unused Advantage Underneath the operational urgency Hirshberg and Shiebler describe, sits an architectural premise about how Artemis is built. In an AI era, both sides draw on the same technology. Whatever edge the defender once held in raw capability is gone. What remains, and what the attacker cannot acquire from outside, is knowledge of the defender's own environment. Who works where. What is normal for this user? Which systems matter to the business? Whether a 3 a.m. login is routine or the first in this person's history. That knowledge has always existed. What has never existed
More from The Security Strategist
AI can flag thousands of suspicious events and create a clear summary of each one. But the bigger question is whether security teams can understand how AI reached its conclusions. In this episode of the Security Strategist Podcast, Dan Williams from Cambridge Intelligence joins John Tolbert, Principal Analyst and Director of Cybersecurity Research at KuppingerCole Analysts, to discuss how AI and visualisation are changing cybersecurity investigations. They look at how graph visualisation, AI and automation can help security teams make sense of complex data. Williams explains how visualisation can turn machine output into something analysts can understand, investigate and act on, while also helping them see what AI is doing behind the scenes. Why Use Graph Visualisation in Security? Graph visualisation maps the links between accounts, devices, IP addresses, malware and vulnerabilities. Analysts can see how an attack connects instead of reading a list of isolated events. Williams traces the idea back to his company's roots. Before cybersecurity, Cambridge Intelligence served law enforcement, anti-fraud teams and intelligence agencies, who have long relied on link analysis. Picture the detective's wall of photos and string. It looks theatrical, but it works because mapping a problem is the natural way to get your head around it. The approach has since been applied well beyond network mapping. Security teams now use graphs to connect threat intelligence, cloud infrastructure, industrial control systems and software bills of materials. Tolbert also points to identity, where graphs can expose links between devices, accounts and potential insider threats. Williams cites a Wiz presentation that described a security graph containing roughly a trillion links. “Everything is a graph, really,” he says. Attackers are looking for a route from A to B, while security teams need to see those connections and understand where they lead. With a trillion links in the picture, that is not something an analyst can work through manually. AI can help sift through that volume, but it raises another question: if the system identifies a connection, can a security team see why it matters? How Does AI Help Investigations? AI is most useful when each approach is doing a specific job. Traditional machine learning can flag anomalies and classify activity, while generative AI can summarise the results and explain them in plain English. Visualisation then gives analysts a way to see what those findings represent. Williams says the industry has moved beyond simply asking where AI can be added. The more practical question is whether the cost can now be justified. How much are organisations spending on tokens, and what are they getting back? He describes this as a move from exploration to explanation. Experienced investigators may still spend hours working through data without knowing exactly what they are looking for. But, he argues, there are many more people who need a clear account of what happened. This now creates another problem because there is too much information. Machine learning can generate a long list of alerts, while large language models can add pages of text to explain them. Neither necessarily makes an incident easier to understand. Williams believes visualisation needs to be part of that process, giving analysts a way to follow the AI's reasoning and see where its findings came from. He compared it to when you use a maps app; you are not trying to examine every road in a city. You have somewhere to go and need to get to your destination. This is what a junior analyst needs: something similar, a way through the relevant data rather than an overview of everything available. It could be a playbook, a timeline or a short interactive narrative. Williams compares the ideal experience to watching a film, where the sequence follows an attacker from the initial entry point through to the exposed data and the misconfiguration that made it possible. If pre
Imagine a customer logged in to your online portal. Their token is valid, and so is their session. Then one number at the end of an API call changes, and the response comes back with a different customer's account. The WAF sees a well-formed request from an authenticated user. The schema check passes. Nothing in the log says attack. That is an abuse of business logic; cybersecurity analyst Jonathan Care refers to it as a series of "perfectly valid calls in a sequence you never anticipated, doing something you never intended." In this episode of The Security Strategist podcast, Jonathan Care, lead analyst at KuppingerCole, sat down with Uri Dorot, Senior Product Marketing Manager at Radware, to talk about possibilities and impossibilities of API tooling, how detection can tell the difference between an attacker and an unusual customer, and the steps a CISO can take when working with a limited budget. Also Watch: Can Runtime Security Keep Autonomous AI Under Control? Key Takeaways Majority of enterprises fail to detect more than 90% of their APIs. Developers are familiar with Shadow APIs, whereas zombie APIs are not. Endpoint tools can be located cheaply via AI tools and crawlers. Attacks based on business logic make use of valid calls; schema checks can’t detect them. Business logic attacks use valid calls, so schema checks miss them. Response inspection catches reused tokens and PII leaving. Identity alone is insufficient; behaviour and context also matter. A gateway does not replace an API security platform. Every discovered API needs a named owner.
In the recent episode of The Security Strategist podcast, host Shubhangi Dua, podcast host and producer at EM360Tech, sits down with Chenxi Wang, Managing General Partner at Rain Capital, a cybersecurity veteran of more than two decades, a Fortune 500 board member and a VentureBeat Women in AI Award winner. They talk about self-improving AI agents (SIA) and their more advanced cousin, recursive self-improving agents (RSIA), from what they are, why they're suddenly dominating research conferences, and why most enterprises, especially regulated ones, aren't remotely ready to run them without a human in the loop. Key Takeaways RSIA updates the mechanism that improves the agent, not just the agent Sandbox "no" feedback pushes agents toward workarounds, not compliance Governance depends on evidence chains and observability, not approval gates Self-improving agents haven't reached production at scale Regulated industries aren't ready for zero-human-in-the-loop agents C-suites should focus on adoption pressure and provable policy enforcement
What happens when the walls of your organisation can no longer keep your data safe? AI, remote work, and fast-moving data are changing how organisations operate, and the traditional perimeter model is becoming harder to maintain. In this episode of the Security Strategist Podcast, host Chris Steffen sits down with Dr Bill Anderson of Mattermost and JP Ayyappan, Director of Product Management at Virtru, to explore why the future of security isn't about stronger walls. It's about protecting the data itself. Zero Trust Beyond the Network Anderson, who trained as a cryptographer, says this blind spot has been around for decades. Organisations built networks like fortresses, assuming that anyone inside could be trusted. But that approach no longer reflects how people work. Employees use phones, home networks and third-party platforms every day, meaning the traditional security perimeter is no longer enough. Major breaches have shown how risky that assumption can be. This is where zero trust comes in. But Anderson argues that many organisations still apply it mainly to networks and devices, rather than to the data moving through them. Collaboration tools have made this gap even clearer. People are creating and sharing sensitive information every day through meetings, chats and shared documents, often without a way to identify and protect that data as soon as it is created. How AI Changes Data Classification The conversation then turned to one of the biggest challenges facing security teams: AI can combine information in ways that are difficult to predict. Anderson described how a series of seemingly harmless questions about flight paths, weapons ranges, and timing could be combined by an AI system to produce sensitive information, even if no one set out to create it. Steffen gave a similar example. A simple first prompt can lead to a series of follow-up questions, with each response shaping the next. By the third or fourth exchange, the AI could produce highly sensitive information without a clear record of how it was created. This creates a problem for traditional data classification. Rules designed for static documents are harder to apply when information is generated and combined in real time. Ayyappan pointed out that AI can also help address this problem. It can be trained to classify and tag information as new context emerges, but only if organisations already have a clear and consistent tagging structure. Security That Protects the Data Rather than relying on network access to determine who can be trusted, Virtru's approach applies permissions directly to the data through an open framework called the Trusted Data Format. Ayyappan explained that files can carry their own access rules, allowing organisations to control who can access them regardless of the network or device being used. This approach keeps protection with the data, even when a file is shared outside the organisation. It can also make it easier to give external partners or emergency teams access to specific information without giving them access to the wider network. AI Agents Need Access Controls Too Both guests agreed that AI agents, integrations and plugins need to be treated as identities within an organisation's access control model. Anderson explained that Mattermost applies the same user, resource and action controls to AI agents as it does to people. This allows organisations to control what each AI system can access. Ayyappan added that two executives using the same AI agent and accessing the same data could receive different answers based on their individual permissions. This allows organisations to use AI without giving every user or system access to all available data. For security teams, both experts recommended starting with clear limits on which AI systems different groups can access. As organisations develop clearer classification rules, more of these controls can be automated. Ayyappan also stressed that data privacy and protection need to be con
Your engineering team assigns an AI agent a seemingly routine overnight task with limited human oversight. By morning, the agent has moved beyond the task’s intended scope. The AI agent spots a weakness in the sandboxed infrastructure. This weakness gives the agent a path to data it is not permitted to access. But your Endpoint Detection and Response (EDR) dashboard doesn’t flag the problem. To the tools in the environment, the AI agent's activity looks like routine processing. That visibility gap is central to the broader question raised by a July 2026 incident, when OpenAI models circumvented sandbox controls and accessed evaluation data hosted on Hugging Face. In this episode of The Security Strategist podcast, host Richard Stiennon, Chief Research Analyst at IT-Harvest, is joined by Brandon Dixon, Co-Founder and CTO of Ent. Together, they unpack how OpenAI models, operating inside an evaluation sandbox with reduced safeguards, circumvented the controls designed to contain them. The conversation moves from that incident into a much bigger question: as AI agents start acting directly on laptops, servers, and enterprise software, what does the control point of the future actually look like? Key Takeaways OpenAI’s Hugging Face incident shows how agents can cross intended boundaries even when individual actions appear permitted. EDR can observe processes and system activity, but it does not inherently understand the context or intent behind them. Agentic tools, remote-control features, ClickFix, FileFix, and malvertising can make risky activity look like normal work. Prompts and tool calls can now provide clues to an agent’s stated objective, but behaviour and context are still needed to determine risk. Built-in agent guardrails provide an important first layer of protection, but they are not a complete security control. For certain workloads, local AI can offer meaningful advantages in cost, speed, privacy, and data sovereignty. Some enterprises are reconsidering endpoint compute and local GPU investments as the economics and risks of cloud-only AI become clearer. Isolated-tenant design, learned at Microsoft, now shapes Ent's architecture. The opportunity is to rebuild endpoint security around context, intent, and real-time prevention.