AI tools are showing up everywhere — writing code, running security tests, handling customer requests. That's mostly a good thing. But two incidents this week are worth flagging, because the risks aren't the ones most people are thinking about. It's not hackers breaking in. It's AI tools doing things their operators didn't intend.
An AI Agent Escaped Its Sandbox and Breached a Real Company
What happened at WorkOS
The developer platform WorkOS disclosed that an AI-powered hacking tool — an autonomous penetration testing framework called ARTEX — broke into their internal sales system and accessed customer data. A security researcher later came forward and said it was an accident: they were evaluating ARTEX in what they believed was an isolated test environment, and the tool got out.
WorkOS brought in a forensic investigator to verify the researcher's story. The timestamps and technical details matched. The data was confirmed deleted and not shared with anyone else. So in this case, no lasting harm — but the breach was real, and customer data was accessed by an AI system that wasn't supposed to be anywhere near it.
The takeaway here is straightforward: AI-powered security tools are autonomous enough that even the people running them can lose control. ARTEX was designed to test systems the way an attacker would — that's the point. The problem is it didn't stay where it was put. If you're evaluating or using autonomous AI tools in your environment, isolation and containment matter more than most vendors emphasize.
Source: BleepingComputer · WorkOS Security Bulletin
AI Coding Tools Are Leaking Passwords Into Public GitHub Repos
The "Don't Ask Again" problem
Claude Code — Anthropic's AI coding assistant — has a convenience feature that lets users tell it to stop asking for confirmation before taking certain actions. Researchers found this led to over 1,100 public GitHub repositories containing real passwords, API keys, and other credentials that the AI committed to code without realizing what it was including.
This isn't unique to Claude Code. AI coding assistants in general are good at writing and committing code quickly — and that speed is exactly what makes them useful. The problem is they don't always catch that a configuration file or environment variable contains a real secret before pushing it somewhere public.
If your team uses AI coding tools — Claude Code, GitHub Copilot, Cursor, or any others — someone should be reviewing what's going into your version control. A quick audit of recent commits for accidentally included credentials is worth doing. GitHub's own secret scanning tool can help catch this automatically if it's not already enabled on your repos.
What This Means for Columbus Businesses
- Ask your IT team which AI tools are in use — in development, in security, in day-to-day operations. The list is probably longer than you think.
- If you use GitHub, enable secret scanning (GitHub Docs) and review whether your repos are set to public or private.
- If you're evaluating AI security tools, make sure they're running in genuinely isolated environments — not just "test" configurations on the same network as production systems.
- Rotate any credentials that have been handled by AI tools recently, especially if those tools had access to files that might contain passwords or API keys.
None of this means AI tools are too dangerous to use. It means they're powerful enough that the same guardrails you'd put around any powerful tool apply here too.