👋 Welcome back to AI for SME Success, your weekly dose of practical AI insights that matter to small businesses.
This week:
- Four recent cases of AI agents deleting data without permission — and the lessons every business owner should take from them.
- Why 70% AI / 30% human is the content ratio that wins now.
- A prompt from Claude Code’s creator to stop AI from going off the rails before it starts.
- Recent AI guides and rollouts across major platforms.
Let’s dive in! 👇
🛑When AI Agents Disobey & Destroy
AI agents have exploded in popularity over the past year, but the guardrails haven’t kept pace. Many business owners don’t realize that AI agents can refuse to follow instructions.
Here is a list of four recent instances when AI agents deleted data without permission. Let’s review what happened and what we can learn.
April 2026, Database and Backups. During a standard task, the Cursor AI agent ran into an issue and responded by wiping the database and three months of backups. The agent went on to write a “confession,” detailing the exact safety rules it had broken. The AI agent didn’t just threaten PocketOS, the Utah-based Software-as-a-Service company where it was deployed. It also put every rental business relying on PocketOS’s platform at risk. (Inc.)
March 2026, Database and Backups. A developer let Claude Code run an open-source “Infrastructure as Code” (IaC) tool on AWS. A missing state file led to duplicate resources, then a “destroy” command that wiped two production sites and every backup snapshot. (Tom’s Hardware)
February 2026, Inbox. The user asked OpenClaw to review an inbox and “suggest what you would delete, but don’t action until I tell you to.” It behaved correctly on a test inbox, but a larger real inbox led to AI memory overflow and loss of instructions. The user noted “I couldn’t stop it from my phone. I had to run to my Mac mini like I was defusing a bomb.” (PC Mag)
November 2025, D Drive. A developer using Google Antigravity asked the AI to delete just the cache. After running the command, the AI agent left the user’s entire D drive completely wiped. (Tom’s Hardware)
Takeaways:
- Store backups where autonomous agents cannot reach
- Avoid destructive words (“delete,” “drop,” “remove”) in production prompts
- Do not run agents with high-level permissions
⚖️ 70/30 AI Content Strategy
The numbers on AI-generated content just got harder to ignore:
- 44% of daily music uploads (Deezer, April 2026)
- 33% of new websites (404 Media, April 2026)
- 21% of videos served to new YouTube users (The Standard, December 2025)
According to Deezer, 97% of users couldn’t tell AI-made music from human tracks in a blind test. However, only 1 to 3% of streams come from AI-generated music, mostly because algorithms detect AI-generated content and remove it from playlists.
The people who really profit from using AI in content generation mass-produce inputs (ideas, drafts, variations), then aggressively test, kill, and amplify only what works.
Takeaways:
Going 100% AI in any content creation is very risky. You may lose trust and algorithms can flag your content as fraudulent. Going 100% human-made is simply not efficient.
70% AI / 30% human-made is the best split, where AI generates drafts, variations, and polished assets at scale so you can test ideas quickly without getting stuck in production.
🚦 Approve AI Plan Before It Acts
I handed AI a workflow and watched it fail twice. First, it burned through tokens, then ran into roadblocks halfway through.
Most AI disasters and disappointments are preventable. You just need to ask AI to show you its plan before it does anything.
Boris Cherny, Anthropic’s technical staff and creator of Claude Code, called this out directly in Mastering Claude Code in 30 minutes.
Before any complex task, prompt: “Brainstorm a step-by-step plan, flag roadblocks, limitations, and options, then wait for my approval before you execute.”
Think of it as Chain-of-Thought reasoning used before AI acts, not after it derails.
🛠️ AI Rollouts Across Major Platforms
Here’s what dropped this week from Anthropic, Google, Microsoft, and OpenAI:
Anthropic
- Claude for Creative Work. Claude’s new positioning as a creative partner, with features aimed at writers, designers, and marketers who don’t have a dedicated creative lead.
- Claude connectors for creative tools. Anthropic releases 9 Claude connectors for creative tools, including Blender and Adobe.
- Canva in Claude Design. Claude now lives inside the Canva design surface.
- Generate files in the Gemini app. Gemini can now produce Docs, Sheets, and Slides from a prompt, closing the artifact gap for Workspace-based teams.
- Gemini Enterprise Agent Platform. Google Cloud offering designed to help businesses build autonomous AI agents.
- Ask Youtube. Chat with any YouTube video. Pull summaries, timestamps, and quotes from any video.
Microsoft
- Copilot in Outlook Agentic drafting, scheduling, and triage built into Outlook.
OpenAI
- Inside GPT-5 for work.A guide from OpenAI shows how businesses apply GPT-5 (including 5.5) improved reasoning and reliability to real workflows.
Thank you for reading today’s edition!
If this issue was valuable, pass it along to a fellow business owner. I’d love to hear your feedback at natalia@nataliabrattan.com.
See you next week,