Recent revelations of artificial intelligence (AI) agents breaking code have further underscored the need for enterprises to bolster their cyber resilience, in particular, skilling up.
As it is, a common challenge enterprises face is being able to patch quickly, said Daniel Aw, Red Hat’s Asia-Pacific vice president and general manager.
He noted that organisations typically would take an average 90 days to deploy a patch, a timeline that was no longer adequate in today’s AI-powered threat landscape.
With the accelerated speed at which vulnerabilities are discovered by AI tools, businesses simply cannot afford to take that long, said Aw, during a media briefing in Singapore.
In fact, the median time-to-exploit for newly identified vulnerabilities dipped to just under one day in 2026, compared to 23 days last year, according to PwC.
Singapore in June had pushed organisations to bolster their cyberdefence, as advancements in frontier AI models posed significant risks to digital systems, in particular, critical information infrastructures.
Highly autonomous and capable of multi-step logic, frontier AI models often are deployed to run complex software codes and advanced analysis, said Singapore’s Cyber Security Agency (CSA).
They can cut the time needed to identify vulnerabilities and create exploits, enabling threat actors to potentially launch critical attacks in mere hours, CSA warned.
A new report from CrowdStrike revealed that threat actors already were using AI agents to run cyberattacks, including an operation during which the agents moved between compromised environments in 12 minutes.
The security vendor said it already detected three times as many agentic intrusions to date, in the past four months alone, compared to the first half of the year.
In one campaign, Chinese-speaking hackers developed multi-agent frameworks running specialised agents in parallel. They accessed Anthropic’s Claude models via a China-based AI proxy service, or transfer station, and built two penetration-testing harnesses that granted Claude Code various levels of autonomy.
One harness provided agentic flexibility that allowed Claude Code to operate as a single agent, either chasing a target based on a standard operating procedure, or delegating actions to multiple specialist agents in parallel.
The hackers also incorporated other alternative models, including DeepSeek and ChatGPT, according to CrowdStrike.
Skilling for AI era
The emergence of AI-assisted attacks and exploits are further compounded by reports of AI agents going rogue and breaching enterprise systems, including Hugging Face, which was targeted in a hacking campaign staged by OpenAI-powered AI agents.
It underscores the need for organisations to build the skillsets required to better manage their ecosystems and a threat landscape that increasingly will be fuelled by AI technologies.
Software libraries, for instance, can be better managed when companies do a stocktake and strip out those they are not utilising, Aw said.
Doing this can narrow the organisation’s surface area, making it more efficient to secure and deploy patches, he said.
He added that Red Hat aims to work with its customers to provide such expertise and skills support through its newly launched Asia-Pacific AI Centre of Excellence in Singapore.
The facility is targeting to hire an initial team of 80, scaling when needed, comprising engineers, architects, and project managers, according to the software vendor.
It noted that the centre will support customers and partners in skills development critical to their own AI deployment, including software engineering, automation, and DevSecOps.
Red Hat added that it would provide its open source technologies and engineering know-how to facilitate organisations looking to develop their AI experiments towards production-ready deployments.
Singapore’s Defence Science and Technology Agency (DSTA) has signed up to work with the US software vendor through the new centre, with plans to explore and co-develop AI and cloud-native applications, which then can be applied to defence solutions.
In their AI deployments, organisations such as DSTA have to focus on decision security across the entire development lifecycle as well as business policies and processes, according to Akash Philip, DSTA’s deputy director of AI development.
Decision security is a framework that secures how critical decisions are made within the enterprise, encompassing DevSecOps workflows, cybersecurity, governance, employees, risk management, and AI engineering.
With defence a diverse sector, where technology, sovereignty, and trust are fundamental, Philip said DSTA’s collaboration with Red Hat will look to tap the latter’s open-source expertise and community outreach.
He added that companies deploying any AI applications, be it agentic or machine learning operations, need the engineering discipline and expertise to identify the right model to use, test it with their intended use cases, and build in the essential security components.
They also have to know where and how they are deploying AI within their ecosystem, he said at the media briefing.
DSTA wants to build the engineering platform to support such requirements, bring in the applications, and facilitate the rigorous testing, he added.
With reports of AI deployments going rogue, such as the OpenAI-Hugging Face incident, AI risk is something companies now are aware of, Philip said.
However, it cannot be a crippler, he said, noting that organisations still want to leverage the technology to improve services for their customers.
They should be cognizant of the risk involved, rather than be crippled by it, and take the necessary measures to manage the risks, such as looking at how and what access AI agents should be given, he added.
