Cybersecurity in the age of frontier AI

The threat landscape has changed overnight

Frontier AI offensive capabilities are not new, but with the emergence of models like Anthropic's Mythos, they are about to become the new normal. In controlled testing, Mythos generated 181 working exploits against Firefox 147 by chaining four distinct bugs, and the median time-to-exploit for newly disclosed vulnerabilities has collapsed from 23 days in 2025 to under one day in 2026.

For boards and security leaders, this is not an incremental shift. It is a structural one. Patch cycles, risk registers, security-testing cadences and incident-response playbooks all assumed defenders had days, sometimes weeks, to act. That assumption no longer holds.

Mythos is a frontier AI model developed by Anthropic and the first it has chosen not to release for general availability, citing concerns that broad release could increase offensive cybersecurity risk if misused. Through Project Glasswing — one of the largest coordinated vulnerability disclosure efforts in the industry's history — approximately 50 vetted organisations, including major launch partners, critical infrastructure providers and open-source maintainers, were given early access to Mythos to scan and patch their own products, backed by $100 million in usage credits.

In pre-release testing, Mythos escaped secured sandboxes, harvested credentials and attempted to conceal unauthorised actions — behaviours that underscored the risks of broad release. Even if Mythos itself never reaches adversaries, comparable capabilities will. Offensive AI is now a baseline assumption, not an edge case.

The volume of known, exploitable vulnerabilities is about to surge beyond what any traditional patch cycle can absorb. Three shifts define the new environment:

1. Exploitation windows are collapsing. Time-to-exploit has dropped from 23 days to under one day. Traditional patch SLAs are materially inadequate.
2. The attack surface is expanding. Every AI agent, plug-in and autonomous workflow is a new identity and a new privilege path. AI agents must be treated as untrusted by default.
3. Patching spikes are coming. Project Glasswing disclosures will be the first of many, with thousands of patches landing across major vendors in the months ahead.

Existing programmes face two bottlenecks: remediating findings at AI speed across the tools you already run and validating that those fixes work without breaking the business. Both are made worse by a real human capacity constraint; even mature security teams are stretched on headcount and specialist skills.

Enterprise technology was not built to manage AI-enabled vulnerability discovery at machine speed. Yet no enterprise wants to scrap the programmes, tools and controls it has built through years of investment. The smart move is to run frontier AI through the programme you already have — extending capability and reach at machine speed, without disrupting or duplicating existing investments.

Through our collaboration with Anthropic, PwC is deploying AI-enabled cyber defence capabilities powered by Claude Opus 4.7, Anthropic's most capable commercially available model. The combination matters: Opus 4.7 delivers the sustained agentic reasoning and production-grade reliability that high-stakes security workflows demand, while PwC brings the integration, operating-model and governance expertise to translate that capability into operational value inside complex enterprise environments. Connected to your existing controls via Model Context Protocol (MCP) and API integration, we deploy:

• Agentic remediation across your current cybersecurity tooling — an AI-enabled harness that determines which defensive levers to pull and activates them.
• Agentic test-case generation integrated with your existing test suites — turning probabilistic model outputs into deterministic test cases, so patches and rule changes can be validated automatically before deployment.
• Pre-production validation orchestration — structured proof, generated before any autonomous change lands, that an action achieved its security objective without unintended business impact.

All of this operates within PwC's autonomy envelope: the decision boundaries, standing authority frameworks and audit trails that let defensive agents act at machine speed on routine actions while humans retain control of scope, escalation thresholds and policy. It is underpinned by PwC's framework for visibility, observability and governance of AI-driven actions, and aligned with Anthropic's safeguards built into Opus 4.7.

The vulnerability onslaught is upon us. The question of whether AI can reshape cyber defence has been answered. What remains is execution — strategic clarity at the board level, operational acceleration across the enterprise, and confidence that AI itself is governed responsibly. The organisations that act now, through the defences they can already trust, will define the new baseline for cyber resilience.