Security in the AI Era
A leadership guide to preparing for cyber risk in the age of AI — and why organizations need to act now.
By Kiran Sahota
The security landscape changed faster than most organizations expected. AI didn't gradually shift the threat environment — it collapsed the timeline. The tools attackers now have access to are the same tools defenders are still learning to use. That asymmetry is the problem.
This is a guide for executive and security leadership on what has changed, what it means, and what to do about it.
The new threat baseline
Three data points define where we are.
The new threat baseline
AI-enabled offensive capability is no longer theoretical.
1,000s
Zero-day vulnerabilities
detected by Anthropic's Mythos model across major technology platforms
73.1%
Exploit success rate
public Claude models detecting and exploiting real vulnerabilities
8 mo.
The capability gap
approximate lag between frontier global models and Chinese counterparts (Deepseek)
Sources: Anthropic Mythos model research (2026); Stanford HAI AI Index 2025
Finding and exploiting software vulnerabilities using AI models is real. Organizations should be prioritizing remediation in months, not years.
Anthropic's Mythos model detected thousands of zero-day vulnerabilities across major technology platforms. Public Claude models already operate at a 73.1% exploit success rate against real vulnerabilities. And the gap between frontier global models and Chinese counterparts like Deepseek is approximately eight months — meaning capability that looks exclusive today becomes widely available by end of year.
The conclusion is direct: finding and exploiting software vulnerabilities using AI is no longer a research problem. Organizations should be measuring remediation timelines in months, not years.
What has changed
The pre-AI security model assumed attackers were constrained by human speed, human cost, and human expertise. A sophisticated attack required skill, time, and resources. That constraint provided a natural rate-limiter on risk.
In the AI era, those constraints are gone. Reconnaissance, vulnerability discovery, and exploit generation can all be accelerated by orders of magnitude. The question is no longer whether your perimeter can resist a determined human attacker — it is whether it can withstand an autonomous AI agent running at machine speed, continuously, at no marginal cost.
The timeline
The window for action is narrow and closing.
The timeline
The window for action is short and getting shorter.
Now
Act as if bad actors already have the agents
Mythos-level capability is currently limited to select partners. Use this window to scan your assets thoroughly and remediate all known vulnerabilities. This lead time will not last.
6–12 months
Public models will close the gap
Frontier advances diffuse rapidly. Patch velocity becomes the new baseline expectation. Organizations without rapid remediation muscle will fall permanently behind.
12–24 months
Autonomous agents at scale
Sophisticated attackers run AI agents continuously. Cyber risk will materially impact organizations that are not prepared. This is the trajectory of current capability curves.
The threat timeline: three phases from today’s narrow window to full AI-agent parity for sophisticated attackers.
The organizations that use the current window well will be in a materially different position when capability reaches full parity. The ones that treat this as a future problem will discover it is a present one.
Four imperatives for security leadership
Against this backdrop, CISOs and executive sponsors need to organize around four disciplines.
Hygiene — see everything you own and patch at machine speed. You cannot protect what you cannot see. Every internet-exposed asset needs to be discovered, inventoried, and continuously monitored. Patching cadences measured in quarters are no longer acceptable. Critical CVEs need to be addressed in days.
Leverage — use AI to defend before bad actors do. The same capabilities that power offensive tools power defensive ones. AI-augmented vulnerability triage, code review, and threat detection are not optional enhancements — they are the only way to match the velocity of AI-powered attacks. Organizations that delay AI adoption in their security function are unilaterally disarming.
Resilience — make it easy to contain. When something goes wrong — and in a high-velocity threat environment, something will — the speed of containment determines the scope of damage. Zero standing privileged access in production, strong segmentation, and tested incident response playbooks are the foundation of resilience.
Governance — treat AI itself as critical infrastructure. AI systems are now operational infrastructure. They need to be inventoried, approved, and monitored the same way any critical system is. An approved-model registry covering 100% of production AI use is not a compliance exercise — it is operational hygiene.
Five pillars with ownership and outcomes
Five pillars — with ownership and outcomes
Each pillar is actionable with clear ownership. Strategy without accountability is a slide deck.
01
Know what you own
CIO + CISO
100% of internet-exposed assets discovered within 24 hours
02
Patch at machine speed
CISO + Head of Engineering
Critical CVEs in production patched within 72 hours, 95% of the time
03
AI-augmented defense
CISO + Head of BC/DR
AI in ≥80% of triage and code-review workflows by Q4
04
Re-think your defenses
CISO + Head of Infrastructure
Zero standing privileged access in production
05
Govern AI itself
CISO + CAIO + General Counsel
Approved-model registry covers 100% of production AI use
Each pillar
has an owner, a measurable outcome, and a timeline
Five pillars of AI-era security — from asset hygiene to AI governance — each with defined ownership and a measurable target outcome.
What to do this quarter
If you are a CISO or executive sponsor, the first move is an honest assessment of where your organization sits across these five pillars. Not where the roadmap says you will be — where you are today.
- Asset inventory. Can you enumerate every internet-exposed asset right now, within 24 hours? If not, that is the first gap to close.
- Patch velocity. What is your current mean time to patch a critical CVE in production? If it is measured in weeks, that number needs to come down before the threat environment catches up.
- AI in defense. Are you using AI tools in your security workflows today? Planning is not enough. Attackers are not planning — they are running.
- Privileged access. Do you have standing privileged access in production? Every account that exists is a potential attack vector.
- AI governance. Do you have a registry of all AI models running in production? If not, you are governing a system you cannot fully see.
The organizations that address these questions methodically over the next two quarters will be in a materially stronger position when the threat environment reaches full parity. The ones that treat it as a future problem will discover it is a present one.
Sources
- Anthropic, Mythos model vulnerability research (2026) — Mythos model detected thousands of zero-day vulnerabilities across major technology platforms, demonstrating AI capability at cybersecurity expert level.
- RAND Corporation, AI and Cybersecurity (2025) — Analysis of AI-enabled offensive cyber capabilities and implications for enterprise security posture.
- CISA, AI Cybersecurity Collaboration Playbook (2025) — Guidance for organizations on integrating AI into defensive security operations.