Your AI Agent Is Only As Secure As The Vendor Running It
Somewhere in your organisation right now, a piece of software is probably making a decision that a human used to make: reordering stock, adjusting a price, approving a routine transaction. The real question isn't whether those decisions are made by AI or automation. It's who authorised the agent, what they're allowed to do, and whether anyone can prove it afterwards.
Third-party risk is no longer just about whether a supplier can access your systems. It's about whether the software representing your business can make decisions safely once it does.
Recent research by Verizon found that third parties were involved in 48% of confirmed breaches, up from 30% the year before and 15% the year before that.
So, how will security play out when AI assistants start taking actions autonomously? The supply chain was already a significant attack path before anyone gave software the authority to spend money.
MoreThan Data
AI raises the stakes for security and trust across consumer-facing purchasing journeys and enterprise operational workflows. Commerce is where this plays out fastest. Agents here can hold standing payment permissions, adjust pricing and inventory as they go, and increasingly deal with other systems instead of people. There may be little time to catch a bad decision once it's already moving.
The early data on how organisations are securing AI systems isn't very encouraging. IBM found that 13% of organisations had already suffered breaches involving AI models or applications, and 97% of those lacked proper AI access controls. Supply chain compromise was the most common cause of AI security incidents, accounting for 30% of cases, including compromised apps, APIs and plug-ins.
The compromise is bad enough, sure. But acting on it is worse. A breached vendor once meant data to contain and disclose. As AI systems become more embedded in commercial operations, the stakes change: a compromised AI agent with ongoing permissions could place orders, move money and alter records before anyone realises something is wrong.
This is where vendor transparency becomes operational security. Businesses can only extend real autonomy to AI systems when they have real visibility into how the underlying platforms are managed, monitored and governed. That visibility can’t be a sales conversation or a marketing claim. It needs to be verifiable, current, and specific enough to survive an audit.
When Annual Audits Aren't Enough
Procurement and security teams need to raise their expectations. A SOC 2 or ISO 27001 certificate is worth having, but it only proves the vendor passed an audit at some point in the past. What you actually want is the detailed report behind it, and you should be able to ask for one.
In practice, security is often reduced to a line on an RFP checklist. That's no longer enough. Vendors should expect to be asked for evidence, not just a yes or no, and the ones that treat transparency as a baseline rather than a differentiator will be the ones enterprises trust with real authority.
The same goes for agent permissions. Handle them like access for a privileged user, kept tight and checked on a schedule rather than set up once and forgotten about. It also matters what the vendor can see when something goes wrong on their end. Does it get logged? Does it trigger a response? Does anyone actually notice? An agent shouldn't come with broad authority by default. That authority should be something a business turns on for a reason, watches and pulls back at will.
Centralised trust centres that offer self-service access to audit reports, penetration test summaries and compliance documentation are a more realistic way to keep up with vendors whose systems and risk profiles continuously change. Regulation is moving in the same direction. The EU's Digital Operational Resilience Act already requires financial institutions to maintain continuous oversight of critical ICT third parties, and that model is unlikely to stay confined to financial services as agentic adoption spreads.
This doesn’t mean putting every vendor under suspicion. Organisations need to understand, however, that autonomy and accountability must scale together. If you give an AI agent the ability to act, you should be just as rigorous about the assurance you demand from the vendor supplying it.
This matters beyond the enterprise too. Trust in AI-driven shopping and operations from the buyer's side will depend on knowing these systems are governed properly.
The businesses that get ahead will not be the ones moving fastest on AI adoption. They’ll be the ones that can prove, on demand, exactly how much trust their vendors have earned.
Dirk Hoerig is Co-founder & Chief Innovation Officer at commercetools
Image: Zach M
You Might Also Read:
A Guide To Taming Agentic AI Risks:
If you like this website and use the comprehensive 8,000-plus service supplier Directory, you can get unrestricted access, including the exclusive in-depth Directors Report series, by signing up for a Premium Subscription.
- Individual £5 per month or £50 per year. Sign Up
- Multi-User, Corporate & Library Accounts Available on Request
- Inquiries: Contact Cyber Security Intelligence
Cyber Security Intelligence: Captured Organised & Accessible






