That risk extends beyond conventional office IT. Government platforms, cloud services and Critical Information Infrastructure (CII) can provide access to essential services and operational-technology environments. If attackers can automate reconnaissance and exploitation across exposed systems, a single overlooked weakness may become easier to locate and abuse at scale.

Frontier models can also strengthen phishing and impersonation campaigns. CSA says AI can generate convincing lures at scale, create realistic voice clones and video deepfakes, and produce tools intended to bypass multi-factor authentication. The result is a broader attack surface: technical vulnerabilities remain important, but identity, permissions and human decision-making become equally attractive targets.

What is known about Claude Mythos Preview

Anthropic has described Claude Mythos Preview as a highly capable model for computer-security tasks and launched Project Glasswing to use it in defensive work on critical software. However, the strongest public claims about the model should be treated cautiously: independent verification remains limited, and Anthropic has kept the model under controlled access rather than releasing it broadly.

Singapore’s Government has said it does not have access to Mythos and does not assume it will receive early access to every future frontier model. Instead, it says it works with AI labs, cybersecurity companies and partners that can track capability developments and assess their safety implications.

That distinction matters. Organisations cannot build their security plan around inspecting one named model. They need controls that remain effective as different models, tools and attack workflows become available.

Singapore’s immediate defensive priorities

CSA and GovTech have alerted CII owners and government agencies to the frontier-AI risk. The immediate measures are deliberately practical:

• Patch exposed systems first. Remediate all critical and high-severity vulnerabilities on internet-facing systems, prioritising assets that could enable widespread compromise.
• Protect administrative access. Enable multi-factor authentication on administrative interfaces, gateways and cloud-management consoles. Where MFA is not technically available, restrict access through controls such as IP allow-listing.
• Review permissions. Check privileged accounts, stale accounts and unnecessary access rights. Least-privilege access and network segmentation can limit what an attacker can reach after an initial compromise.
• Improve visibility. Maintain an accurate inventory of critical systems, internet-facing assets, cloud services and third-party dependencies, then monitor them continuously.
• Prepare to respond. Incident-response plans should include clear escalation paths, tested backups and exercises that reflect faster, more automated attacks.

These steps are not novel, but frontier AI makes delay more dangerous. The goal is to reduce the number of weaknesses available to attackers and shorten the time between detection and remediation.

Use AI to defend—but secure the AI itself

AI-enabled cybersecurity can help teams detect threats, triage alerts, investigate incidents and identify or remediate vulnerabilities more quickly. Singapore’s CSIT describes AI-driven workflows and fuzzing systems designed to uncover zero-day vulnerabilities in mission-critical software, alongside adversarial emulation to expose architectural weaknesses before attackers find them.

Defensive AI still creates its own security responsibilities. CSA says hardening the model alone is insufficient. System owners should assess risks throughout the AI lifecycle: planning and design, development, deployment, operations and maintenance, and end of life. That means protecting training and operational data, interfaces, dependencies, supply chains, access controls and monitoring—not only the model’s behaviour.

Human validation remains essential, particularly when an AI system recommends a patch, labels an incident or proposes a response that could affect operational technology or public services.

A shared response for CII owners and businesses

Singapore’s policy direction combines baseline security with continuous testing. CII owners already use third-party audits and regular penetration tests, with reports submitted to CSA; the Government has also emphasised the need to strengthen controls continuously rather than rely only on periodic validation.

The operating principle is succinct: lock down, find first, fix fast. In practice, that means maintaining asset inventories, scanning continuously, patching rapidly, segmenting networks and privileges, testing software and dependencies, and rehearsing recovery. Businesses that are not designated CII owners should apply the same priorities to internet-facing services, identity systems, cloud consoles and sensitive third-party connections.

Collaboration and threat intelligence

No single organisation can observe every frontier-AI capability or every attack path. Government agencies, CII operators, cloud and technology providers, security researchers and academia should share actionable indicators, vulnerability information and lessons from exercises as quickly as possible.

Some intelligence will necessarily remain classified. Singapore has said it will selectively share classified threat intelligence with CII owners while also equipping them with defensive tools. Organisations should therefore act on CSA advisories and use sanitised, shareable intelligence where detailed sources cannot be distributed.

National exercises are part of that model. The Critical Infrastructure Defence Exercise has used AI-driven attack simulations across 11 critical sectors with industry partners to build resilience and improve coordination.

The practical takeaway

Frontier AI does not replace the fundamentals of cybersecurity; it raises the cost of neglecting them. Singapore’s strongest near-term defence is a combination of rapid patching, MFA, least-privilege access, continuous monitoring, adversarial testing, secure AI adoption and fast information sharing. The organisations that find their weaknesses before an automated attacker does will have the greatest advantage as frontier-model capabilities continue to develop.

That risk extends beyond conventional office IT. Government platforms, cloud services and Critical Information Infrastructure (CII) can provide access to essential services and operational-technology environments. If attackers can automate reconnaissance and exploitation across exposed systems, a single overlooked weakness may become easier to locate and abuse at scale.

Frontier models can also strengthen phishing and impersonation campaigns. CSA says AI can generate convincing lures at scale, create realistic voice clones and video deepfakes, and produce tools intended to bypass multi-factor authentication. The result is a broader attack surface: technical vulnerabilities remain important, but identity, permissions and human decision-making become equally attractive targets.

What is known about Claude Mythos Preview

Anthropic has described Claude Mythos Preview as a highly capable model for computer-security tasks and launched Project Glasswing to use it in defensive work on critical software. However, the strongest public claims about the model should be treated cautiously: independent verification remains limited, and Anthropic has kept the model under controlled access rather than releasing it broadly.

Singapore’s Government has said it does not have access to Mythos and does not assume it will receive early access to every future frontier model. Instead, it says it works with AI labs, cybersecurity companies and partners that can track capability developments and assess their safety implications.

That distinction matters. Organisations cannot build their security plan around inspecting one named model. They need controls that remain effective as different models, tools and attack workflows become available.

Singapore’s immediate defensive priorities

CSA and GovTech have alerted CII owners and government agencies to the frontier-AI risk. The immediate measures are deliberately practical:

• Patch exposed systems first. Remediate all critical and high-severity vulnerabilities on internet-facing systems, prioritising assets that could enable widespread compromise.
• Protect administrative access. Enable multi-factor authentication on administrative interfaces, gateways and cloud-management consoles. Where MFA is not technically available, restrict access through controls such as IP allow-listing.
• Review permissions. Check privileged accounts, stale accounts and unnecessary access rights. Least-privilege access and network segmentation can limit what an attacker can reach after an initial compromise.
• Improve visibility. Maintain an accurate inventory of critical systems, internet-facing assets, cloud services and third-party dependencies, then monitor them continuously.
• Prepare to respond. Incident-response plans should include clear escalation paths, tested backups and exercises that reflect faster, more automated attacks.

These steps are not novel, but frontier AI makes delay more dangerous. The goal is to reduce the number of weaknesses available to attackers and shorten the time between detection and remediation.

Use AI to defend—but secure the AI itself

AI-enabled cybersecurity can help teams detect threats, triage alerts, investigate incidents and identify or remediate vulnerabilities more quickly. Singapore’s CSIT describes AI-driven workflows and fuzzing systems designed to uncover zero-day vulnerabilities in mission-critical software, alongside adversarial emulation to expose architectural weaknesses before attackers find them.

Defensive AI still creates its own security responsibilities. CSA says hardening the model alone is insufficient. System owners should assess risks throughout the AI lifecycle: planning and design, development, deployment, operations and maintenance, and end of life. That means protecting training and operational data, interfaces, dependencies, supply chains, access controls and monitoring—not only the model’s behaviour.

Human validation remains essential, particularly when an AI system recommends a patch, labels an incident or proposes a response that could affect operational technology or public services.

A shared response for CII owners and businesses

Singapore’s policy direction combines baseline security with continuous testing. CII owners already use third-party audits and regular penetration tests, with reports submitted to CSA; the Government has also emphasised the need to strengthen controls continuously rather than rely only on periodic validation.

The operating principle is succinct: lock down, find first, fix fast. In practice, that means maintaining asset inventories, scanning continuously, patching rapidly, segmenting networks and privileges, testing software and dependencies, and rehearsing recovery. Businesses that are not designated CII owners should apply the same priorities to internet-facing services, identity systems, cloud consoles and sensitive third-party connections.

Collaboration and threat intelligence

No single organisation can observe every frontier-AI capability or every attack path. Government agencies, CII operators, cloud and technology providers, security researchers and academia should share actionable indicators, vulnerability information and lessons from exercises as quickly as possible.

Some intelligence will necessarily remain classified. Singapore has said it will selectively share classified threat intelligence with CII owners while also equipping them with defensive tools. Organisations should therefore act on CSA advisories and use sanitised, shareable intelligence where detailed sources cannot be distributed.

National exercises are part of that model. The Critical Infrastructure Defence Exercise has used AI-driven attack simulations across 11 critical sectors with industry partners to build resilience and improve coordination.

The practical takeaway

Frontier AI does not replace the fundamentals of cybersecurity; it raises the cost of neglecting them. Singapore’s strongest near-term defence is a combination of rapid patching, MFA, least-privilege access, continuous monitoring, adversarial testing, secure AI adoption and fast information sharing. The organisations that find their weaknesses before an automated attacker does will have the greatest advantage as frontier-model capabilities continue to develop.