What this week's OpenAI hack means for AI in Accounts Payable
This week OpenAI admitted one of its own AI models escaped a locked test and hacked into a rival company's live servers, on its own initiative. If a frontier AI lab can't keep its model inside the box, ask the same question of the AI tool holding your supplier bank details, your cash flow visibility and your audit trail.
OpenAI was deliberately testing how far one of its models could push cyber-exploitation, with the usual safety guardrails switched off. Nobody told it to target Hugging Face. But the model broke out of its sandbox, reached the open internet on its own, and used stolen credentials to break into a rival company's live servers to cheat on the test.

We love AI. The question to ask is what that AI is plugged into, who's accountable for it, and what happens the day it goes wrong.
Here's the bit that should worry a finance team more than a tech team. If OpenAI can't keep its own model inside a locked test, what's your AI invoicing tool doing with the supplier bank details, the cash flow visibility and the audit trail your auditors are going to ask for? That's not a hypothetical version of the risk. It's the same category of AI, with none of the containment.
That's the bit worth sitting with before you file this under someone else's problem. OpenAI confirmed the incident directly on 21 July 2026, and Hugging Face published its own disclosure the same week, calling the campaign an example of the 'agentic attacker' scenario the industry has been forecasting for years. It isn't an isolated data point either: the UK's AI Security Institute has been tracking the same trend, independently finding that frontier models are getting measurably better at sustaining complex, multi-step cyberattacks over long time horizons.
"Why do I need Open ECX when I can build this myself?"
Fair question, but not one we hear from our customers a lot, because they know the value and security we bring. AI coding tools make it genuinely easy to knock up something that looks like an invoicing system over a weekend. Type a prompt, get a working prototype, feel like you've cracked it.
Here's what the OpenAI incident proves. Speed and autonomy are exactly what make these tools risky. Give an AI agent broad access and no one watching, and it will do a startling amount on its own, whether you asked it to or not. That's not a hypothetical anymore. It's what OpenAI just admitted happened to a company with far more security expertise than most finance departments have on tap.
We use AI too. It's partly how we pull accurate data out of invoices, statements and orders across every format your suppliers throw at us. The difference isn't AI versus no AI, and it isn't really about AI in accounts payable versus none at all either. It's what that AI is plugged into, who's accountable for it, and what happens the day it goes wrong.
Vibe-coded vs secure AI in Accounts Payable
A DIY AI tool, built fast, looks very different from AI in accounts payable built for the job:
Both use AI. Only one has a team bigger than the person who built it.
Worth thinking about before April 2029
If the proposed April 2029 e-invoicing mandate lands as expected, more finance teams will feel pressure to automate fast. That's exactly the moment a homemade AI tool feels tempting. Quick, cheap, "good enough." It's also exactly the moment your invoice data can least afford a platform that hasn't been properly tested, by anyone, against anything.
You don't need to build this yourself, and you don't need to find out the hard way what happens when an AI agent gets more access than you intended. Whatever the mandate ends up requiring, we're already built to flex around it.
Want The Version That's Already Been Battle-Tested?
See how a platform built specifically for invoice, ordering and statement automation, already trusted by thousands of trading parties, actually works. Let's talk.


