AI agents changed the security math, and most businesses haven't adjusted
Your systems now include agents that hold keys, touch data and act on their own. That's a new attack surface. The threats, what still works, and the practical list.
Your business has a new kind of employee, and attackers noticed
By most 2026 estimates, more than 90% of companies are running autonomous AI agents somewhere: customer-bot, code assistant, data pipeline. Each of those agents is a new identity inside your network, holding an API key, a credential or a data connection. Cloud Security Alliance research now puts the ratio of non-human identities to human ones at around 45 to 1, and those machine identities are exactly what attackers are learning to hunt.
This isn't theoretical. Security vendors tracking the shift report 65% of enterprises responding to a 2026 survey had suffered at least one AI-agent security incident in the previous year, and nearly a third were willing to concede the agent became fully compromised. Prompt injection, crafting input that tricks an agent into acting outside its instructions, now ranks as the top failure vector in agentic systems.
The threats that are new
None of these replaces classic threats, phishing, misconfiguration, ransomware are all alive and well. They layer on top, which is precisely the problem: a frontier where an agent's judgement, not just a human's click, becomes the thing that gets breached.
- Prompt injection, feeding an agent (via an email, a file or a web page) instructions that bypass its guardrails. It's the number one vector today
- Vibe coding risk, AI-generated code is everywhere, and Trend Micro research found roughly 45% of vibe-coded modules carry security flaws by default
- Leaked AI secrets, 28.65M hardcoded secrets were added to public GitHub repositories in 2025, with AI-related secrets up 81% year-on-year
- Deepfake voice and video, synthetic media targeting finance teams and supplier verification is now cheap enough for anyone to run
- Shadow AI, employee use of unapproved AI tools tripled to 45% by 2025 (Verizon DBIR), creating blind spots in who touches what data
What still works
For all the novelty, the hard-won basics are the same and they fix most of it. The agents and the code that produced them still run on the same infrastructure you already know how to lock down.
- Keys and secrets out of code, into a vault, never hardcoded, always rotated
- Least privilege. An agent gets exactly the scopes its task needs, nothing else
- Strong authentication everywhere, passkeys over passwords, MFA on human accounts
- Review AI-generated code before it ships. Treat 'the model wrote it' as a red flag, not a pass
- Watch the integrations your team quietly spins up, shadow AI is a data-exfiltration path
The practical list
If you run software or services that touch customer data, the same checklists apply with one addition: know what your agents can see and do, and how you'd know if someone started talking to them sideways. Reviewing MFA, rotating keys and auditing AI you've adopted is a weekend project, an incident audit is a quarter of a bad year.
Put this to work in your business
The posts here come from what we see building and running systems every day. If any of it maps to your business, we can help make it operational, AI projects, web platforms, data and security included.
