Singapore’s CSA Warns AI Could Shrink Cyberattack Timelines from Months to Hours

That speed changes the defensive equation. A vulnerability that once gave an organisation weeks or months to detect and remediate could potentially be found and weaponised much sooner. Traditional patching cycles, manual investigation and periodic security reviews may no longer be fast enough for exposed systems.

Anthropic reported that Mythos Preview was capable of identifying and exploiting zero-day vulnerabilities in every major operating system and major web browser when directed by a user . That claim describes the model’s tested capability, not evidence that it had been used in live attacks.

Risks CSA highlighted

Internet-facing systems become higher-value targets

Unpatched critical and high-severity vulnerabilities on internet-facing assets were the most urgent concern. These systems are directly reachable from outside an organisation and could offer attackers a rapidly discoverable entry point .

CSA’s warning applies beyond large technology companies. Banks, operators of critical information infrastructure, cloud users and any business exposing applications, administrative interfaces or development environments to the public internet could face greater pressure to find and fix weaknesses quickly.

Autonomous discovery lowers the barrier to exploitation

The significance of Mythos Preview is the reported combination of vulnerability discovery and exploit generation. Anthropic said its testing showed the model could move from finding a previously unknown flaw to producing a working exploit across major operating systems and browsers .

That could make sophisticated offensive techniques more accessible and allow attackers to examine more targets in less time. It also raises the risk that several individually minor weaknesses could be combined into a broader compromise, although the available CSA reporting does not establish that such misuse has occurred.

Financial institutions face amplified systemic risk

Singapore’s financial sector treated the issue as an amplification of existing cyber risk. The Monetary Authority of Singapore convened bank chief executives to discuss the implications of frontier models capable of autonomously identifying and exploiting vulnerabilities .

The practical implication is that organisations should not wait for a wholly new threat category or a confirmed local incident. Existing weaknesses—especially exposed, unpatched and over-privileged systems—may become more dangerous when attackers can search for them at AI-assisted speed.

Immediate measures CSA urged

CSA’s short-term guidance focused on reducing the easiest and most consequential attack paths:

• Patch every critical and high-severity vulnerability on internet-facing systems. These assets have the greatest exposure to automated attacks and the potential for widespread impact if compromised .
• Require multi-factor authentication for administrative interfaces, gateways and cloud management consoles. Where MFA cannot be used, organisations should apply alternative restrictions such as IP allowlisting .
• Remove unnecessary public access to development and staging environments. Secure them with strict access controls or disconnect them from the internet where possible .
• Review cloud configurations, with particular attention to publicly accessible resources and overly permissive security groups .
• Apply least privilege. Remove unnecessary permissions, revoke dormant accounts and limit service-account and API access to what is required .
• Enable and correctly configure DDoS protection for internet-facing assets .

These steps are deliberately familiar. Their urgency comes from the possibility that AI reduces the time between exposure, discovery and exploitation.

Longer-term changes to security architecture

CSA’s broader recommendations point to a shift from periodic hardening toward continuous exposure management and layered defence.

Reduce the attack surface

Organisations should remove unnecessary internet-facing services, disable unused ports and protocols, and correct insecure cloud settings. The fewer reachable systems an automated attacker can map, the fewer opportunities there are for rapid exploitation .

Limit lateral movement

Network segmentation and micro-segmentation can help contain an intrusion. If an attacker gains an initial foothold, separating systems and restricting east-west traffic can make it harder to reach sensitive services or privileged infrastructure .

Treat suppliers and dependencies as part of the exposure

AI-assisted discovery does not stop at an organisation’s own code. Dependency scanning, controlled update processes and vendor-risk assessments can help identify weaknesses in third-party components before they become an indirect route into critical systems .

Monitor continuously for attack paths and abnormal behaviour

CSA urged organisations to improve visibility across network traffic, endpoints and user activity, especially around privileged accounts and lateral movement. Behavioural analytics and anomaly detection may help identify attacks that progress faster than conventional, human-operated campaigns .

Build security into applications and identity systems

The longer-term approach also includes defence in depth, zero-trust principles, continuous verification, runtime application security monitoring and stronger identity-and-access management. Mandatory MFA, rapid credential-response processes and continuous authentication can reduce the damage caused by compromised accounts .

What Singapore did next

On 5 May 2026, CSA wrote to the boards and senior leadership of Critical Information Infrastructure owners, asking them to conduct a fresh review of cyber risks that specifically accounts for AI-enabled threats . The review was intended to examine areas including critical systems, internet-facing assets, privileged access, cloud services and third-party dependencies .

The government also said it did not have access to Mythos and was not aware of any local bank that had been granted access. It said it was working with partners that did have access to assess the model’s implications . At the same time, Singapore said it was accelerating its own use of AI for defensive cybersecurity, including automated vulnerability and patch testing .

What organisations should take from the warning

The CSA advisory does not establish that Mythos Preview has been used in an active cyberattack. Its message is about preparedness: the defensive window may become shorter even if the underlying vulnerabilities and attack techniques remain familiar.

For most organisations, the first response is therefore practical rather than experimental: identify exposed assets, patch the highest-risk flaws, remove unnecessary access, secure cloud environments and ensure that detection and response can operate continuously. Frontier AI makes those fundamentals more time-sensitive—not less important.