Six months after Mythos: is your organisation Frontier Ready?

On 7 April 2026, Anthropic announced Claude Mythos Preview, an AI model that found thousands of previously unknown vulnerabilities across every major operating system and browser, and wrote working exploits for many of them with little human help. Rather than release it, Anthropic gave early access to a group of major software vendors and open source maintainers through Project Glasswing, so they could fix their own products first.

Six months on, the headlines have moved on. The problem has not. This month the Cloud Security Alliance published the Frontier Ready Maturity Model, a framework for measuring how well an organisation can defend against attackers using AI of this kind. It is a good moment to look back at what organisations could have done since April, and to take an honest measure of where yours stands.

April to October: what happened

The industry responded quickly. Within a week of the announcement, the CSA’s CISO community, SANS, [un]prompted and the OWASP GenAI Security Project published an expedited briefing, Building a “Mythos-ready” Security Program. It set out eleven priority actions on an aggressive timetable: some to start that week, others within 45 days, 90 days and six months.

The vulnerabilities were real, and there were many of them. In late May, Anthropic reported that Glasswing partners had found more than 10,000 high- or critical-severity vulnerabilities in widely used software in about seven weeks. Independent firms reviewing a sample confirmed more than 90% as genuine.

Fixing, not finding, became the bottleneck. Of the first 1,596 vulnerabilities Anthropic disclosed to open source projects, fewer than 100 were known to be patched by late May. Every one of those fixes eventually becomes an update that organisations have to test and deploy.

Australian guidance pointed back to the fundamentals. The Australian Signals Directorate updated its guidance on frontier AI at the end of April. In late May, Australian Government agencies were told to get the basics right before reaching for frontier AI themselves, with a focus on Essential Eight Maturity Level Two for patching and application hardening. You do not need access to the most advanced models to defend well.

And the capability is spreading. Anthropic said in May that it expected Mythos-level models to become widely available within six to twelve months. We are now inside that window.

It is worth keeping one balancing point in mind. Faster exploitation has not yet produced a matching rise in damage, and most serious incidents still start with stolen credentials, social engineering or a compromised supplier. The fundamentals still matter most. What has changed is how quickly they need to work.

What a well-prepared organisation did between April and now

Looking back over the six months, the organisations in the best position took a handful of practical steps. None of them required a large new budget.

• April: they told the board. Leadership was briefed, and risk reporting stopped assuming there would always be weeks to patch.
• May: they found out what faces the internet. They built or refreshed an inventory of internet-facing systems, gave every system an owner, and agreed a fast-track path for critical fixes so they did not wait for the monthly window.
• June and July: they closed the obvious doors. Phishing-resistant MFA for every administrator, a review of standing admin rights in Microsoft 365 and Entra ID, servers stopped from freely calling out to the internet, and the most important systems separated from everything else.
• August and September: they rehearsed. They checked backups by actually restoring them, practised handling several serious incidents in the same week, and agreed which containment actions, such as isolating a laptop or disabling an account, could happen without waiting for approval.
• Throughout: they looked at their own AI. They found out which AI tools and agents their people were using and what those agents could reach.

If that sounds like your last six months, you are in good shape. If it does not, you are in good company, and the new framework gives you a way to start.

The Frontier Ready Maturity Model, in brief

The model’s premise is simple: frontier AI did not invent new ways in. It shortened the clock on the existing ones, multiplied the number of attackers who can use them, and raised the skill of the least capable. Being too small or too obscure to be noticed is no longer a defence.

It measures twelve areas, grouped into three domains:

Domain	Areas
The programme: know and govern	Governance and risk · Exposure management · Defensive AI and agent operations · Workforce and operating model
The planes: the zero trust surface	Network segmentation and egress · Identity and access · Endpoint and workload hardening · Telemetry and deception
The flows: absorb and recover	Pipeline and supply chain security · Vulnerability operations · Security operations · Resilience and recovery

Each area is scored from Level 1 to Level 5. What separates the levels is not how much paperwork exists, but tempo: how quickly the capability works, and whether it keeps working when several things go wrong at once.

• Level 1: running on a pre-AI calendar, such as monthly patching or quarterly scans, or not in place at all.
• Level 2: the gap is visible. The capability exists and is measured, even if it does not run continuously.
• Level 3: continuous operation at human pace, with named owners and thresholds.
• Level 4: machine-speed operation that holds under concurrent incidents, with people setting the rules rather than approving every action.
• Level 5: self-improving, and adapting faster than the threat.

Two features make it particularly practical.

It sets a realistic target. The framework recommends that most organisations aim for Level 3 across the board, and Level 4 where speed matters most: vulnerability operations, security operations, revoking compromised access, and recovery. Level 5 everywhere is rarely the right goal.

It recognises that many organisations rely on providers. There is a separate profile for organisations that lean on managed service providers rather than a large in-house team. The same outcomes are measured either way; what changes is the evidence and cadence. One principle holds for both: even when monitoring is outsourced, the organisation should own its logs, their retention and its records.

Five areas to check first

The framework’s authors recommend a bias towards action over documentation. If you only look at five areas this quarter, we would start here:

1. Vulnerability operations. Are you fixing vulnerabilities as fast as new ones arrive? The framework’s headline measure is simple: findings resolved divided by findings created. Below one, your backlog is growing. Internet-facing critical fixes should have a path to deployment within 24 hours.
2. Identity and access. Do administrators sign in only with phishing-resistant credentials? Is admin access granted just in time rather than held permanently? Could you revoke every token and session of a compromised account or integration in minutes?
3. Exposure management. Do you know everything that faces the internet, who owns it, and which few systems would genuinely hurt to lose?
4. Security operations. How long does it take from a credible alert to a laptop being isolated or an account being disabled? Have you measured it, even if the number is uncomfortable?
5. Resilience and recovery. When did you last restore a critical system from backup, and could you rotate its credentials and rebuild it in a known order?

The framework suggests a 30-minute qualitative self-assessment as a first step: read each area’s row and decide honestly where you sit. Most organisations find it more useful than a long audit.

Where Summit fits

Summit Cyber Group helps Australian organisations move up the model where it matters:

• Exposure management: Tenable-powered discovery, prioritisation and remediation tracking, run as a continuous programme. Summit is a Tenable Assure Gold Partner.
• Cloud and identity security reviews: Microsoft 365, Entra ID, Azure and AWS configuration, privileged access and phishing-resistant MFA.
• SIEM engineering: Splunk, Microsoft Sentinel and Cribl pipelines and detection content built for faster triage, with logs that stay under your control.
• Managed security services: firewall, email, endpoint, identity and SIEM, monitored around the clock.
• Managed AI security: visibility and control over the AI tools and agents your people use.
• Web application security assessments: testing your applications before someone else does.

Perth-based, senior people. If you would like to work through the Frontier Ready self-assessment and agree realistic targets for each area, talk to us.

Sources: Cloud Security Alliance, “Frontier Ready Maturity Model” (public edition, version 1.0, October 2026); CSA CISO Community, SANS, [un]prompted and OWASP GenAI Security Project, “The AI Vulnerability Storm: Building a ‘Mythos-ready’ Security Program” (version 1.0, May 2026); Anthropic, Project Glasswing update (May 2026); Australian Signals Directorate and Australian Government guidance on frontier AI (2026).