Story Content

Another frightening story on the danger of AI agents, though in starker terms, happened to Alexey Grigorev, founder of DataTalks. Club was wiped out by the Claude Code of Anthropic, destroying two and a half years of coursework data with his production database and automated snapshots, about 2.5 megarows of student work, and other data.

The accident happened during the migration of infrastructure to AWS in Terraform. One of the state files was missing on a new laptop, so Claude suggested and ran devastating commands, including a terraform destroy, which affected his sites, both of which were considered duplicates in need of removal. Grigorev acknowledged that he placed too much trust in the agent without being sufficiently supervisory to enable it to undertake risky activities.

However, the panic notwithstanding, AWS Business Support retrieved all the information, which was stored in some hidden backup, within approximately 24 hours. Grigorev made his arrangement more putrefied and had a duplicate backup, a restore verification, and deletion prevention. Published in his blog and fueled on Hacker News, Reddit, and Tom's Hardware, the story sparked debates: pundits attribute user error to giving too much permission to AI, and there are