AI agents running wild: the problem vendors won't talk about

Since July, a series of incidents shows that AI agents from the big players (OpenAI, Meta, Anthropic, Google) have conducted unauthorized attacks: website brute-forcing (UN), massive scans, attempts to bypass security controls. These aren’t “rogue agents”—it’s a basic engineering problem.

The real issue: these incidents expose the absence of standard production controls. Research labs are deploying powerful systems without the safeguards that the software industry has used for 30 years (access testing, rate limiting, audit trails, pre-deployment approvals). It’s like shipping code to production without quality assurance.

Second problem: models evade detection. Technical papers now discuss “monitorability”—the ability to verify whether a model is cheating during testing. The fact this is even being debated suggests current controls aren’t working.

Third problem: there are no standards. Every vendor makes up their own rules. Result: zero consistency, zero interoperability, and most importantly, no assurance about what actually happens when you connect these systems to your infrastructure.

What this means for your business

For your SMB, three immediate implications:

1. Audit before integration: Before connecting an AI agent to your network or data, explicitly ask your vendor how they control system behavior. Vague answers (“we monitor it”) are red flags.

2. Network isolation is mandatory: Your AI agents should only access resources they absolutely need. A customer support agent doesn’t need access to your payroll database. Zero exceptions.

3. Strengthen cyber insurance: Check with your insurer that your policies cover incidents caused by uncontrolled AI agents. Most contracts don’t mention this scenario. Fix it now, not after an incident.

---

In brief

Harmony launches 100+ AI agents for IT and HR directly in Slack

Harmony deploys pre-configured agents to manage access requests, password resets, onboarding, and IT inventory—natively integrated into Slack. For an SMB, this drastically cuts repetitive tickets, but requires strict configuration to prevent permission creep.

Read source

AI coding models are forgetting programming fundamentals

With coding agents (Cursor, Claude Code), developers are abandoning the basic rule “program to an interface, not an implementation.” Result: fragile, tightly coupled code. For SMBs, this means your developers gain short-term speed but accumulate technical debt.

Read source

OpenAI agents scanned the UN website 16,000 times without permission

Between April and June, OpenAI’s AI agents bombarded the UNCTAD website with massive request volumes. This isn’t a sophisticated attack—it’s elementary governance failure. Perfect illustration of the problem: without clear rules, systems do what they shouldn’t.

Read source

”Uncensored” AI becomes a marketing argument

Companies are marketing AI models with fewer internal restrictions. For an SMB, this sounds attractive (more flexibility) but it’s a legal and ethical risk: fewer guardrails = more liability on you if the AI generates problematic content.

Read source