The AI blind spot: what boards see versus what's actually happening

Anthropic released Claude Mythos Preview on 7 April through Project Glasswing, granting controlled access to selected partners rather than making the model public. The reported capabilities were strong enough to trigger rapid responses from supervisors, policymakers and security institutions across Europe, the UK and the US.

The European Central Bank began probing institutions on model-driven intrusion scenarios, including incident playbooks, mean time to patch, dependency depth and privilege escalation frequency. In the UK, the Cross Market Operational Resilience Group (CMORG) convened the Bank of England, FCA, HM Treasury and NCSC around one table.

The time between vulnerability discovery, attack readiness and organisational response is collapsing.

Cyber exposure is now forming across tools, workflows, providers, data flows, infrastructure and decisions — often beyond the systems leadership reviews. Boards and executive teams need a new baseline: where exposure forms, how fast it moves, who can detect it, who can contain it, and who can prove control.

Why speed matters now

Anthropic claimed Mythos identified thousands of zero-day vulnerabilities across major operating systems and browsers, including flaws hidden for decades. Mythos is also able to connect vulnerabilities into working exploit sequences — including examples that would take elite human researchers days or weeks.

Most companies still manage cyber risk through human-speed structures: committees, policies, vendor reviews, annual assessments, escalation protocols, incident playbooks. AI-enabled exposure appears in code, credentials, suppliers, workflows, employee behaviour, model access and data movement.

• 14h — expert task completed by AI in a simulated 32-step network takeover
• $10B — estimated global damage from NotPetya in 2017, the historical parallel
• 3wk — from Mythos technical release to regulatory escalation across three jurisdictions

The business consequence

NZZ draws a useful historical parallel with NotPetya, the 2017 cyberattack that hit Maersk, Merck, Mondelez and many other organisations. Maersk's operations across 76 ports and nearly 800 ships were disrupted. Its own damage was estimated at at least $300M, while global damage was estimated at around $10B.

A software vulnerability can become operational paralysis. Mythos raises the next business concern: vulnerability discovery, testing and attack preparation are becoming faster, cheaper and more repeatable. Cyber moves from technical risk to response-time risk.

The blind spot

1. Leaders see approved tools. Field reality: usage sits outside the official map

The dashboard says adoption is progressing. The policy says usage is controlled. The roadmap says capability is scaling. Yet the most valuable AI use often sits with teams building their own apps and workflows, professionals choosing tools that perform better, employees combining approved and unapproved tools, and business units adapting faster than central governance.

The visibility gap is the danger and it is called Shadow AI. The organisation may benefit from AI activity it cannot fully see, validate or secure. Leadership measures formal adoption. Risk forms in real usage.

2. Leaders see vendors. Field reality: dependency embeds below visibility

Vendors are approved. Cloud exposure is understood. Third-party risk is monitored. Procurement has reviewed the contracts. Security has assessed the platforms. Yet AI creates deeper dependency chains: model providers, cloud infrastructure, APIs, embedded AI features inside software platforms, data flows across jurisdictions.

A tool used for convenience can become infrastructure. A feature can become a dependency. A vendor can become difficult to replace. Project Glasswing also creates an access asymmetry: a coalition of 40–50 named and unnamed organisations, primarily US-based technology groups and selected partners, receiving controlled access before broader disclosure.

3. Leaders see monitored risk. Field reality: exposure is moving faster than monitoring

The AI Safety UK conducted the only public evaluation of Mythos. In a simulated 32-step takeover of an undefended corporate network — a task estimated to take a human expert roughly 14 hours — Mythos completed the test. Discovery accelerates. Exposure surfaces faster. Response remains organisational.

4. Leaders see cyber as technical. Field reality: the easiest entry point is human

Credentials remain one of the simplest routes in: attackers do not always need to break in when they can log in. AI makes phishing, social engineering and targeted deception more convincing, multilingual and personalised. This brings AI security risk closer to every organisation.

The response-time gap

The standard rises. Detection must become faster. Containment must become more automated. Response must become more rehearsed. Ownership must become clearer. AI-side speed runs in minutes, hours and days. Most organisations still respond in days, weeks, quarters and annual cycles.

Fears of AI agents successfully attacking well-protected systems on their own remain exaggerated for now. That caveat sharpens the leadership point: the gap between well-defended and poorly mapped organisations becomes more consequential.

What boards should expect to be asked

Supervisors are already asking about operational readiness: incident playbooks, mean time to patch, dependency depth, privilege escalation and model-driven intrusion scenarios. Boards should prepare for harder evidence-of-control questions:

1. How fast can we patch critical vulnerabilities?
2. Which AI-related providers are embedded in critical workflows?
3. Can we detect abnormal behaviour automatically?
4. Can we isolate accounts or systems within seconds?
5. Have we tested incident playbooks against AI-enabled scenarios?
6. Which decisions or workflows are AI-assisted but not traceable?
7. Who owns risk when AI capability sits outside the official stack?

These questions belong in the boardroom because they define resilience.

The boardroom debate

Most organisations are active on AI. They have policies. They have pilots. They have approved tools. They have governance committees. They have risk frameworks. Yet many are governing AI through an outdated version of the organisation. The field has already moved.

Workflows have changed. Dependencies have deepened. Security exposure has accelerated. Decision chains have become harder to trace. The technical layer is only the entry point. The structural exposure sits in workflows, dependencies, accountability and response time.

The most dangerous AI risk may already be operating inside everyday work: useful, productive, decentralised, embedded — and not fully visible.

The next phase of leadership will be judged on evidence of control.

Article by Barbara Cresti.