The biggest shift has already happened. AI isn't an add-on to cybersecurity — it's the core

For years, AI in cybersecurity was described as an "accelerator" — a tool to help an analyst sift through more logs or write rules faster. That framing is now obsolete. In 2026, AI stopped being a defensive flourish and became the foundation of both attack and defense. The pivot point is behind us.

At Visera PSA we watch this from up close, building tooling for Obsigen AI — an agent designed for Red Team operations and the work of security teams. Below is what's really changed, separating the hard facts from the narrative that's been spinning recently.

Models can now find and weaponise vulnerabilities on their own

The clearest evidence of a paradigm shift is Claude Mythos Preview from Anthropic — a model capable enough in cyber that the company chose not to release it publicly. Instead it went into a restricted programme called Project Glasswing, where dozens of partners (including AWS, Apple, Google, Microsoft, NVIDIA, Cisco, CrowdStrike, Palo Alto Networks, JPMorganChase and the Linux Foundation) use it to harden critical software.

What can a model like this actually do? According to Anthropic's published results, Mythos autonomously:

• maps code and the attack surface,
• chains several individual weaknesses (in one case two to four low-severity bugs) into a working privilege-escalation chain,
• generates and tests exploits without human involvement after the initial prompt.

The headline numbers are real. Mythos found a 27-year-old vulnerability in OpenBSD (a system whose entire reputation rests on security) and a 16-year-old bug in FFmpeg's H.264 decoder that had survived over five million fuzzing runs. It also built an exploit for the wolfSSL cryptographic library (CVE-2026-5194), capable of forging TLS certificates and impersonating, say, a bank. That isn't marketing — those are confirmed, patched vulnerabilities.

The honest cost story matters too, because narratives here often drift. Anthropic's full research campaign cost on the order of $20,000, while the single run that happened to land on the OpenBSD flaw came in under $50. The catch: you can't know in advance which run will be the hit — and that's exactly what changes the economics.

Asymmetry grows — but stick to the numbers

It's often said that the time from disclosure to "weaponisation" has collapsed from years to minutes. The direction is correct, but the specific numbers are worth checking before they end up in a board deck.

The VulnCheck State of Exploitation 2026 report shows that roughly 29% of known exploited vulnerabilities (KEVs) in 2025 were attacked on or before the CVE publication date — in practice, as zero-days. CrowdStrike's 2026 Global Threat Report gives a higher figure: 42% of exploited vulnerabilities were attacked before public disclosure. The median time to first exploit dropped below five days, and adversaries keep getting faster.

Those numbers genuinely should worry you — and you don't need to round them up to "67%" to make the point. The conclusion is clear: the response window has shrunk to days, sometimes hours. A defender has to protect everything; an attacker only needs one way in.

A good recent illustration: the "Copy Fail" vulnerability in the Linux kernel (CVE-2026-31431), sitting in the code for nine years, was found with AI assistance in about an hour, and a working exploit fitted into 732 bytes of Python. It shows how dramatically the entry bar has moved.

Open-source loses "security through transparency"

For a long time the assumption was that open code is safer because "many eyes" review it. AI flips that logic. The same many eyes can now belong to an automated scanner that combs through an entire repository in minutes.

The best example is AISLE, which in January 2026 was responsible for finding all 12 vulnerabilities in a single OpenSSL security release — a library considered one of the most thoroughly audited in the world. Some of those bugs went all the way back to 1998. Code transparency stops being a guarantee of safety; it becomes a vector for powerful scanners.

The other side of the coin: AI as defender and as Red Team

The same technology that lowers the bar for attack radically reinforces defense. A modern SOC powered by language models analyses logs, prioritises alerts and runs automated playbooks (SOAR) without manual intervention. XDR platforms increasingly understand application and infrastructure context, not just signatures.

The most important change concerns testing. Once an attacker has an autonomous agent, a defense based on an annual pentest stops being enough. That's where we see the role of Obsigen AI — an agent that runs continuous, autonomous Red Team operations: it maps the attack surface, tests chains of vulnerabilities and simulates real campaigns at the tempo today's adversaries actually operate. It's exactly the mindset the regulator is forcing through TLPT (Threat-Led Penetration Testing) under the DORA regulation that now binds the EU financial sector.

What it means for your organisation — practically

The paradigm shift doesn't demand a one-day revolution, but it does require deliberate decisions in a few areas.

1. Weave AI through the whole application lifecycle. From threat modelling, through AI-assisted code review, to AI-driven fuzzing and continuous resilience tests. The earlier a flaw is found by your team (rather than by someone outside), the cheaper it is.

2. Build governance for AI agents. Every model and agent needs an owner, an unambiguous identifier and least-privilege access (ideally just-in-time). Monitor behaviour — deviations from the baseline (e.g. unauthorised exploit generation) must be detectable. And don't forget "shadow AI" — tooling used outside security's line of sight.

3. Move from occasional to continuous testing. A classic once-a-year pentest can't keep up with an adversary operating in hours. Autonomous penetration tests — whether as PTaaS or through an agent like Obsigen AI — give you a current picture of resilience, not a twelve-month-old snapshot.

4. Think about sovereignty and regulation. The absence of European players among Project Glasswing founders is sparking debate about EU technological dependence. In parallel, the EU AI Act (with key obligations for high-risk systems coming into force in 2026) and DORA require auditability for decisions taken by AI. It's worth knowing today whose models are processing your data and how you'll document their behaviour.

5. Educate the whole team, not just security. AI risk is now a topic for every employee who touches code or data, not only the SOC.

The future: AI as a shared language of offense and defense

There are no more "cyber-specialised models" — every next large language model will have offensive and defensive capabilities built in. The security stack will get more "AI-first", and regulatory frames will force an audit trail for automated decisions. The advantage will go to those who treat AI not as a gadget but as a foundation — and build it into the whole security lifecycle.

The biggest shift has already happened. The question isn't whether your organisation will start using AI in cyber defense; it's whether you'll do so before someone on the other side does. At Visera PSA we build Obsigen AI precisely so that answer is on your side.

All numerical data in this article come from publicly available sources (Anthropic / Project Glasswing, AISLE, VulnCheck State of Exploitation 2026, CrowdStrike 2026 Global Threat Report, Theori, among others). Where figures circulating online were inflated, we used the verified values.