The author launched the site ihavebeenclawed.com — an index of 58 documented cases where coding AI agents and chatbots deleted data, disclosed secrets, spent money, or created promises that had to be kept. The list includes 37 affected agents and services, including Claude Code, Codex, Cursor, Gemini CLI, Copilot, Cline, and Replit, and about 90% of the incidents are marked as preventable.

image

What happened

The project was presented publicly through a Show HN post on Hacker News. Each entry in the index has a link to the original source and a highlighted “lesson,” and the “preventable” tag is present in about 90% of the cases. Among the documented examples: on August 5, 2026, Claude Opus 5 in Claude Code created a requested backup in the wrong directory and then launched a recursive deletion across the disk — the severity of the incident was rated 5 out of 5. On May 18, 2026, Cursor, due to broken quote escaping in a path with spaces, recursively deleted a significant part of a Windows user profile, including Desktop and Documents, bypassing the Recycle Bin.

Context

These incidents should be understood in the context of the fact that modern coding agents have shell access, meaning the right to execute commands in the user's file system, and auto-execution modes allow them to work without final human review. The failure class is the same across different vendors: recursive deletions after errors in paths and escaping, overwriting entire configs instead of targeted edits, continuing to work after a failed file operation, and auto-executing destructive commands without confirmation. ihavebeenclawed.com itself is a content project without a business model: the author manually collects publicly described cases, and its value lies in the fact that the same scenarios are documented across dozens of products, not in a single isolated failure.

Why this matters for the industry

For the industry, this is effectively a free corpus of real failures of agentic products: the same classes of failures repeat across different vendors, including the developers of Claude Code, Cursor, Codex, and Gemini CLI, so the index is suitable as a basis for security checklists for agentic products and as an argument in favor of sandboxes, write-scope restrictions, snapshots, and confirmation of destructive commands. It is also useful for eval methodology: paths with spaces and quotes, recursive deletion, overwriting an entire config instead of a targeted patch, and continuing to work after a failed file operation — these are ready-made reproducible scenarios for red-teaming and regression testing. If the index continues to be populated with the same classes of failures, pressure on vendors to adopt such protective features will likely increase, but this is an interpretation, not an established fact.

Why this matters for users

If you use Claude Code, Cursor, Codex, Gemini CLI, or Cline, the index collects real data-loss scenarios with specific measures that can be applied immediately and do not require waiting for vendor fixes: keep backups and large data outside the agent's write zone, commit the project to git before the first request, disable auto-execution for destructive operations, manually check paths before recursive deletion, and distinguish between “suggest” and “execute” modes in the agent's permissions. The case of the Windows profile deleted bypassing the Recycle Bin shows that the risk concerns not only code in the project but also personal files on the same machine.

What is still unknown / limitations

The index is not a measurement of failure frequency, but a curated corpus with sample bias: it only includes documented and publicly described cases. Without a denominator, that is, the total number of agent sessions, no claims can be made about how often agents lose user data. The “about 90% preventable” tag is the author's assessment, not an independent audit. In addition, the project is new: the Show HN post had minimal response on Hacker News at the time of the material's preparation, so the scale of the discussion and further population of the index are currently unclear.

Sources

Author

Look at AI, editorial team