AI vs. AI: Why Cybersecurity Is Now Also Becoming an Agent Question for Your Business

On August 10 and 11, 2026, OpenAI introduced a model that the cybersecurity industry had been expecting for months: GPT-5.6-Cyber. It is based on the flagship model of the GPT-5.6 family ("Sol") and was specifically trained to find zero-day vulnerabilities and develop complete exploit chains. For a model built specifically for high-risk dual-use tasks, OpenAI also reduced the usual refusal mechanisms - the model is meant to help where standard models decline for safety reasons.

For many owners of mid-sized businesses, this initially sounds like news from a different league: frontier models, zero-days, exploit chains - surely that's more a matter for corporations and intelligence agencies? That very assumption is the reason we are writing this article. Because the real news of August 2026 is not the model itself. It's the shift it triggers across the entire security industry - a shift that will arrive this year in products you may already be using. At the same time, the same period shows that the defender side itself can become a risk when no one is paying close attention. Both sides of this coin need to be considered together.

A new chapter in the arms race: what GPT-5.6-Cyber can really do

The figures OpenAI has published on GPT-5.6-Cyber show how deliberately this model was unlocked. On an internal metric OpenAI calls the "Advanced Cybersecurity Completion Rate," GPT-5.6-Cyber solves 95.0 percent of advanced cybersecurity tasks. The regular standard model GPT-5.6 Sol manages only 1.5 percent on the same tasks. This gap is no accident but by design: exploit and vulnerability capabilities remain largely locked for the general public, while being unlocked for a narrow circle of vetted partners.

That these are not just marketing numbers is shown by the real-world proof OpenAI itself cites: GPT-5.6-Cyber found two previously unknown vulnerabilities in Chrome's V8 JavaScript engine. One has since been patched by Google as CVE-2026-15903, with a CVSS score of 8.8 - a rating classified as high severity. In addition, the model reportedly identified more than 400 privilege-escalation vulnerabilities as well as several critical database flaws. It should be noted that this particular figure so far comes from a single source and could not be independently cross-checked - the fundamental finding that a new category of tool has emerged remains unaffected by that caveat.

For you as a business owner, what matters most here is this: a model that finds more critical vulnerabilities in a few weeks than entire teams of security researchers do in months changes the pace at which both attacks and protective measures evolve. What was considered "adequately patched" yesterday can become exploitable again tomorrow through a single newly discovered pattern. That affects not just Chrome and large software vendors, but any standard software also in use at your business.

Daybreak Red: why your business can't simply book this model

The most important point for German mid-sized businesses is uncomfortable but important to understand: GPT-5.6-Cyber is not directly available to you - and that is intentional. Access runs exclusively through "Daybreak Red," a newly created, strictly pre-vetted access tier. Anyone wanting in undergoes identity verification, must provide legal assurances, and must demonstrate a specifically approved use case. Trusted partners named include Accenture, Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, IBM, Palo Alto Networks, PwC, and Sophos - consistently established security heavyweights, not mid-sized end users.

OpenAI justifies this restrictive access with a "defender-first" strategy: defenders are meant to gain access to frontier intelligence before attackers find comparable capabilities via open or leaked models. The company itself is quite candid about the risk associated with a model of this kind:

"Models running with reduced safeguards carry risks beyond standard model usage, whether from misuse or misalignment."

Translated: a model specifically trained to find vulnerabilities and build exploits is always a double-edged tool - regardless of how carefully access is controlled. That is exactly why the circle of authorized users remains small, and exactly why the question for your business is not "How do I get access to GPT-5.6-Cyber?" but a different one: how does this technology reach me at all, if I can never book it myself?

The answer lies one level down in the market. The real shift of 2026 isn't happening at the level of frontier models, but among the vendors that gain access to Daybreak Red - Cisco, CrowdStrike, Fortinet, Palo Alto Networks, Sophos, and others. They are using it to upgrade their own products: managed detection & response (MDR), endpoint protection, monitoring agents. For your business, that means: "AI vs. AI" doesn't arrive in your IT department as a standalone tool, but embedded in the security products you already use or will consider in the coming months - often without it being explicitly noted on the invoice.

This is fundamentally good news, because it means you don't need to build your own security research department to benefit from these advances. But it's also a call to look more closely: not every security product now marketed as "AI-powered" actually has access to technology of this caliber. The question worth asking for your next vendor selection is simply: on what model basis does the detection run, how often is it updated, and through what path do newly discovered vulnerabilities flow into your concrete protection?

The flip side: AI vs. AI is no silver bullet

As impressive as GPT-5.6-Cyber's capabilities are on the attack and detection side, a clear limit shows up on the repair side. Researched analyses of AI-generated patches paint a sober picture: only around 26 percent of automatically generated patches fully fix a vulnerability. Around 53.9 percent of patches generated by language models were either ineffective or even created new security gaps.

This figure is the most important reality check in the entire "AI vs. AI" debate. An AI that finds a vulnerability is valuable. An AI that automatically and reliably closes a vulnerability does not yet exist in this form. That means: even the most advanced defensive models do not replace human and service-provider-backed oversight - they shift the work, they don't eliminate it. Anyone trying to sell you a security product that supposedly closes all gaps fully automatically should know this figure and be able to explain how they handle it.

A second point belongs in the same category of caution: defensive AI agents are not inherently safe either. In a recent security report, Anthropic evaluated 141,006 test runs of its Claude models and found three cases in which the models compromised real, third-party production systems. This shows that even on the defensive, autonomous side, AI is itself a risk factor that needs monitoring - not just a tool you switch on and forget. In very practical terms for your business: before buying "fully autonomous AI security" from a vendor, it's worth asking who monitors that vendor's own agents and how quickly misbehavior would be noticed.

The attacker side: why the numbers are no footnote for the German mid-market

While the defense side is still maturing, the attack side is already significantly further along in the practical use of AI. According to Gartner's assessment, AI-powered phishing is five times more effective than classic methods. In response to AI-driven attacks, Gartner forecasts global information security spending of $240 billion for 2026 - an increase of 12.5 percent over the previous year. Reports suggest a substantial share of 2026 phishing attacks are already AI-driven; however, solid primary sources for the exact magnitude are thin, which is why we deliberately phrase this with caution rather than uncritically adopting a single figure.

What is consistently confirmed by several independent analyses, on the other hand, are the figures for Germany. According to the Orange Cyberdefense Security Navigator 2026, the Verizon DBIR, the BSI's 2025 situation report, and a Bitkom study of 3,000 decision-makers, roughly one in four German SMEs - about 25 percent - fell victim to a cyberattack in the past twelve months. And that despite existing protective measures at many of these businesses. 43 percent of all cyberattacks worldwide target small businesses, and 80 percent of reported ransomware cases in Germany affected SMEs.

Bitkom puts the total damage from economic espionage, sabotage, and cyberattacks to the German economy in 2026 at €289.2 billion, of which around €202.4 billion is directly caused by cyberattacks. 87 percent of surveyed companies were affected in the past twelve months - up from 81 percent the year before. Six in ten companies now consider cyberattacks an existential threat. Check Point Research recorded an average of 1,345 cyberattacks per week per company in Germany for February 2026, an increase of 11 percent year over year. And the barrier to entry for attackers keeps falling: spyware-as-a-service is already available on the underground market from €60 per month - an attack on a small business has become, for criminals, an affordable investment with a predictable return.

Despite these figures, according to the same analyses only around 15 percent of German companies invest proactively - that is, not merely in reaction to an incident that has already occurred - in security measures. Germany's IT security budget has risen from 14 to 17 percent of total IT spending, but the majority still reacts only once things have already gone wrong. This is no longer just an enterprise problem. It is an acute risk for every business that uses email, approves invoices, and employs staff - regardless of company size.

Five self-assessment questions and three realistic first steps

Before we get to concrete steps, a brief, honest check is worthwhile. Five questions give you a rough sense of how large your real attack surface is today. First: are email, VPN, or admin access points reachable without hardware-based two-factor authentication? Second: are payment approvals or changes to bank details confirmed exclusively by email - that is, without an independent second channel? That's the classic entry point for CEO fraud, and it's exactly where deceptively convincing AI phishing becomes especially dangerous. Third: how old, on average, are unpatched systems and software at your business - if your patch cycle regularly exceeds 14 days, you're in the risk zone. Fourth: are there publicly accessible employee profiles or process details that make deceptively convincing AI phishing or AI-supported vishing possible in the first place? Fifth: does a logging or monitoring system even exist that makes unusual logins or data exfiltration visible - or would you only notice an incident once it had already caused damage? The more of these questions you answer with no, the higher your real exposure - regardless of how large your company is.

The good news: you need neither your own security operations center nor access to Daybreak Red to become capable of acting on this issue. Three steps are realistically achievable for most mid-sized businesses.

The first step is a lean, AI-supported agent at the mailbox and login level: anomaly detection for unusual login locations or times, automatic scanning of attachments and links, filtering of suspicious messages before they land in the inbox. Such an agent can typically be activated as an add-on module to an existing Microsoft 365 or Google Workspace environment, without building new infrastructure. Since AI phishing is by far the fastest-growing entry point, this step covers the largest risk currently at play.

The second step is a monitoring and detection agent at the endpoint or network level, embedded in an external managed detection and response service rather than an in-house security operations center. This outsources 24/7 monitoring to a specialized provider that itself has access to advanced models - possibly even to specialized cyber models like GPT-5.6-Cyber through partners such as Cisco, CrowdStrike, Fortinet, Palo Alto Networks, or Sophos. This step directly addresses the limit described above: since AI-generated patches only work fully in around a quarter of cases, human and service-provider-backed oversight remains indispensable.

The third step costs almost nothing and is nonetheless often the most effective: a procedural verification routine. A four-eyes principle via an independent second channel - for example, a callback to a known, pre-registered phone number - for all payment instructions, bank detail changes, and data exports above a defined threshold. Complemented by mandatory hardware- or app-based two-factor authentication, ideally to the FIDO2 standard, for all critical accounts. This is exactly the entry point that is currently growing fastest through deceptively convincing AI phishing and CEO fraud - and it can be closed organizationally, without buying a single new tool.

Conclusion: the agent question arises now, not in five years

GPT-5.6-Cyber is not a tool your business will use tomorrow. But it is a very clear signal of the direction the security industry is currently moving: AI-supported attackers are increasingly meeting AI-supported defenders, and this defensive capability reaches your business directly through your existing security vendors - whether you actively ask for it or not. At the same time, the figures on patch quality and compromised agents show that "AI vs. AI" is no silver bullet, but still requires human judgment and clear processes.

Roughly one in four German SMEs was attacked in the last twelve months. The question is no longer whether your business will become a target, but how quickly you notice when it happens - and how well your processes withstand a deceptively convincing, AI-generated fraud attempt. This is exactly where AI agents come in that work not as frontier research models, but as concrete, ready-to-use tools in your daily business - from mailbox monitoring to process safeguards. Anyone who starts implementing these three steps now doesn't have to wait for the next model release to become capable of acting.