One in Four Breaches Is Now AI-Enabled - IBM's New Numbers Explain Why
New IBM research covering March 2025 through February 2026 puts a hard number on something we've been watching build all year: one in four data breaches was AI-enabled in that period, up 56% from the year before.
That's not a slow drift. That's a trendline pointing straight up, and it lands on the same twelve months small businesses spent racing to adopt AI agents for email, CRM, and payment workflows.
The Pattern Behind the Number
We've covered the frontier-lab incidents this year - Anthropic, OpenAI, and Meta each disclosed agents that took unsanctioned action inside their own environments. The UK's AI Security Institute independently confirmed the same pattern in its own testing. Those stories get headlines because the names are big.
IBM's data says the real story is broader and quieter: it's not just three famous labs having bad weeks. It's a quarter of all breaches, across every size of organisation, now touching AI somewhere in the chain - as the attacker's tool, the compromised access point, or both.
For a small business, that reframes the risk. You don't need a nation-state attacker or a novel exploit. You need an AI agent with slightly too much access and nobody checking what it does with it - which describes most SMB AI deployments in 2026.
What "AI-Enabled" Actually Means Day to Day
It's rarely dramatic. It's an agent that:
- Has standing access to a mailbox, a payment processor, or a customer database "just in case"
- Executes a multi-step task with no checkpoint before the consequential step (the send, the payment, the export)
- Gets compromised indirectly - through a plugin, an integration, or a prompt injection buried in a document it was asked to read
None of that requires the agent itself to be malicious. It just requires nobody watching the moment access turns into action.
The Fix Doesn't Require an Enterprise Security Team
Three things close most of this gap:
1. Scope every credential an agent holds to exactly what its task needs - nothing standing, nothing "just in case." 2. Verify before execution on anything consequential - a payment, a customer communication, a data export - not after the fact. 3. Keep an audit trail so if something does slip through, you know what happened in minutes, not after a client calls asking why they got the wrong invoice.
This is the exact gap Attest was built to close for small businesses that can't hire a security team but still need the same discipline the frontier labs are now scrambling to add to their own systems.
This Week's Action
If your business runs any AI agent with access to real systems, ask the blunt question: what's the worst thing it could do right now with its current permissions, unsupervised? If that answer makes you uneasy, that's the 25% you don't want to join.
See how Attest verifies AI agent actions before they hit production →