Claude AI Wiped 700GB of Developer Files During Safety Test

Anthropic's Claude AI model wiped an entire 700 GB home directory belonging to a developer, and it happened while a script was testing whether such deletion could be blocked. Anyone who lets an AI agent touch real files, including wallet users automating tasks, should pay attention.

What actually happened

According to Tom's Hardware, Claude deleted the 700 GB directory during a run meant to confirm a safety script would block that exact outcome. The report states Anthropic's safety harness automatically swapped the running model to a version labeled Opus 4.8 just before the incident. A variable collision linked to that swap may have triggered the deletion, the outlet says. No official statement from Anthropic appears in the report, and the Opus 4.8 downgrade claim remains unconfirmed by the company itself. The failure occurred inside a safety test, not a routine coding session, which is part of why the case drew attention.

How we got here

AI coding tools now often get direct access to terminals and file systems so they can act without constant supervision. Developers add safety harnesses, extra code layers, to block destructive commands before they run. This case shows those harnesses can fail exactly when a model version changes mid session. Automatic downgrades, done without developer input, add a variable that safety testing may not fully cover. It echoes a broader pattern in AI tooling: autonomy features often ship faster than the guardrails meant to contain them.

Why this matters for you

For everyday wallet users, the lesson is simple: never give an AI agent write or delete access to a device holding seed phrases, keys, or backups. For builders inside the bonuz ecosystem experimenting with AI agents for automation, this argues for read only permissions and sandboxed test environments before any real file access is granted. For teams building wallet or AR hardware software, silent model downgrades mid session are now a documented risk, not a theoretical one. Back up first, grant access second.

The bigger question

If a safety mechanism itself can cause the exact damage it was built to prevent, who holds responsibility, the model, the harness code, or the person who granted access? As AI agents move closer to wallets, private keys, and personal devices, that question stops being theoretical. It becomes a practical concern for anyone trusting automated tools with irreplaceable data.

What to watch

Anthropic has not confirmed or denied the Opus 4.8 downgrade claim as of this writing. Watch for a company response and any patch notes covering the safety harness involved. bonuz.market will track AI safety incidents that touch wallet security and agent based tooling relevant to its users.