The last few weeks have had a pattern. Every week, another company announces an AI agent, an AI assistant or an AI workflow. Every week, another headline tells us something has been breached, leaked or drained. Those two streams of news are connected.
On 30 October, Binance suffered the biggest theft in crypto history, with US$6.4 billion gone, and Chainalysis’s reconstruction says the outcome was decided in 54 minutes. Binance is a well-funded exchange with a large security team, so this was not a case of a small shop forgetting to lock the door. The speed is the lesson. Attackers no longer move at human pace, and many of them now use agents to do it.
That matters well beyond crypto. A recent study found that AI agents broke into 100 firms for pocket change, and that Southeast Asia’s small online shops fit the target profile. For years, obscurity protected smaller businesses, because breaking in took skill and patience. Agents have made it cheap, and cheap attacks get aimed at everyone.
Also Read: The hidden risk in AI adoption: Unchecked agent privileges
The background rate is climbing as well. Global organisations faced an average of 2,336 cyber attacks per week in July, up 16 per cent from a year earlier, with education, energy and travel among the sectors under growing pressure. Supply chains add to the exposure: in Southeast Asia, supply chain attacks are becoming routine, in a region short of security talent and heavily dependent on third-party technology providers. Add agents that can probe thousands of targets cheaply, and the economics tilt further toward the attacker.
We are building faster than we are securing
The picture inside companies is not reassuring. Gravitee’s research found that most organisations have deployed AI agents into production, yet fewer than one in five have full security approval in place. Adobe’s Southeast Asia findings point the other way, with 57 per cent of firms stuck in AI’s “frozen middle”, unable to get projects past their own IT and security teams. Some firms are rushing ahead unsecured while others are stuck waiting on security. Neither is a healthy way to adopt a technology.
The risk grows once agents start acting on their own. An OutSystems report found that enterprise executives are confident they can keep their AI agents in line, but the room goes quiet when they are asked what happens if two agents disagree. Elsewhere, technology leaders in the region are learning to trust AI agents, but not with production. That caution is sensible, and it is also an admission: the controls most firms rely on were designed for people, not for software that makes decisions.
The other half of the problem is us. Many incidents begin with something ordinary: a reused password, a convincing email, or an employee pasting a supplier contract into a free chatbot to save time. That is what shadow AI looks like in practice, and banning AI does not fix it. The phishing playbook is also changing. Check Point Research’s latest data shows phishers now want you to type the malware in yourself, and AI-written lures are eroding trust in the old cues staff were trained to spot.
Security spend should match AI spend
My view is simple. If we are investing seriously in AI, we should invest at least as seriously in securing it, and I would argue for more. Every agent we deploy is one more actor with access to data and systems. As AmanTrue’s Evin McMullen puts it, AI agents need someone standing behind them. That means identity, accountability and limits on what each agent can touch.
Over the past four weeks we have met companies approaching this from different directions. Three stood out: Dropbox, Yubico and Aleph. Each answers a different part of the problem, and I would encourage you to look closely at all three.
Some are building guardrails around the data itself. Dropbox, which has been careful to stay neutral in the ChatGPT, Claude and Gemini contest, has launched Dropbox Protect, and it deserves attention from anyone putting AI on top of company files. It is less about classic cyber defence and more about the question every AI rollout should answer first: what happens to our data when people run large language models over it? Protect, formerly Nira, shows who can access company files across Dropbox, Google Drive and Microsoft 365, so overshared content can be found and fixed before AI tools surface it. If your documents already live in Dropbox, have that conversation now, before shadow AI forces it on you.
Others are fixing access, and this is where I would start. Phishing, reused passwords and stolen credentials all end the same way: someone logs in who should not. A hardware security key such as Yubico’s YubiKey makes that far harder, because it is phishing-resistant and a stolen password alone is no longer enough. Yubico told us that some customers place orders of around 100,000 keys at a time, partly because of staff turnover and keys that are never handed back. That is the scale at which organisations are already treating access as a hardware problem. I believe a large share of today’s incidents could be prevented by securing access this way, and the same discipline should extend to the agents acting on our behalf.
Also Read: Singapore leads on security governance but struggles to enforce it, report finds
And a new generation of firms is going after segments the big vendors have underserved. Scouter’s entry into Singapore’s crowded cyber market is a recent example.
Aleph, launched on 13 October by Red Alpha Singapore and nuKudo, goes after the people side of the problem. It is a technology workforce provider whose portfolio spans AI, cybersecurity, cloud and DevOps, and it hires on aptitude rather than CVs. With security talent in short supply, that matters: you cannot buy protection you do not have the people to run. If your team is stretched, Aleph is a partner worth calling.
What a security twin looks like in practice
A security twin is a simple rule: every AI budget line gets a matching security line, approved at the same time and tracked by the same owner. It is not a fixed ratio, and I would not pretend one number fits every company. It is a discipline that stops security from being the step that blocks a launch or the afterthought that follows a breach.
In practice, I would start with four questions. Which AI tools and agents are running in the company today, including the ones nobody approved? Who or what is each one acting as, and what can it reach? Where does our data go when it is used with a model? And who is accountable when something goes wrong?
The last question matters most. UOB’s research finds that banks want AI agents to move money, as long as someone is accountable, and Penta’s report makes the wider point that cyber incidents are now about trust and leadership as much as firewalls. Boards will not accept “the agent did it” as an answer.
Each partner above maps to one of those questions. Dropbox Protect covers where data goes. Hardware keys from Yubico cover who or what gets in. Aleph covers the people to run all of it. The rest is governance, which costs less than a breach and, as security reports keep showing, often fails at the point of follow-through rather than intent.
The three partners at a glance
Dropbox Protect is a data access governance product. It gives admins one view of who can access files across Dropbox, Google Drive and Microsoft 365, shows which files are shared internally, externally or publicly, and flags overshared files with their owners. Teams can remove public or company-wide links in bulk, revoke access for external users and former employees, and transfer ownership. Policies can alert on or automatically fix risky sharing, such as confidential files shared with personal accounts, and an exportable log supports audits and access reviews. For AI, the point is to restrict overshared content before AI tools and large language models can surface it.
Yubico makes hardware security keys in which the private key is generated on the device and never leaves it, which is what makes them phishing-resistant. The range includes the YubiKey 5 Series, a FIPS-validated series for regulated sectors, the YubiKey Bio Series, the Security Key Series and the YubiHSM hardware security module. The keys support FIDO2, passkeys and PKI, work without a phone or network connection, and integrate with identity providers including Microsoft Entra ID, Okta, Ping and Google Workspace. YubiKey as a Service covers pre-registered keys, backups and global fulfilment for large rollouts. In May, OpenAI added an opt-in setting for ChatGPT and Codex that removes password sign-in and lets users choose a security key.
Also Read: Singapore’s AI adoption surges, but data complexity raises security risks: Report
Aleph is a technology workforce solutions provider formed by Red Alpha Singapore and nuKudo, launched on 13 October. Where a traditional recruiter places candidates, Aleph combines talent assessment, training, deployment, professional services and technology solutions in one model. Its offerings include Select, Train, Deploy, Assessment as a Service, an In-House Talent Engine, Staffing Plus, Custom Workforce Solutions and AI workforce transformation. The roles it covers span AI, cybersecurity, cloud and DevOps, software and data engineering and network infrastructure, including Forward Deployed Engineers. Candidates are assessed on aptitude and problem-solving rather than CVs.
What we are doing, and what we ask of you
At e27, we are running a Startup Clinic for a curated group of startups from the Cybersecurity Research Programme Office (CRPO) technology translation programmes. These are teams turning research into products, and the market needs more of them. We would like the wider ecosystem to meet them at the Demo Day: investors looking for early exposure to security, and enterprises willing to pilot technology built close to home.
If you build, buy or back technology, here is the ask. Treat security as part of your AI plan from day one. Put guardrails around your data before AI touches it, which is the case for Dropbox Protect. Secure access with hardware keys such as Yubico’s, so a stolen password is no longer enough. Close the talent gap with partners such as Aleph, rather than waiting for the market to produce more security engineers. And watch this space. In an AI world that changes every quarter, it is becoming one of the most important ones.
The post Equal spend: why AI budgets need a security twin appeared first on e27.