[Claude Opus 4.6 (1M context)] Claude deletes files without user permission and repeatedly assumes instead of asking

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude modified files I didn't ask it to modify

What You Asked Claude to Do

What You Asked Claude to Do:

I was working with Claude on a coaching homework document (iterating through drafts).
During the conversation I said "the other post: move to @to-make-content folder" —
referring to a file Claude had generated earlier. Instead of moving it, Claude had
already deleted it moments earlier without being asked. When I pointed out I never
asked for deletion, Claude acknowledged the mistake but the file was already gone.

What Claude Did Instead:

1. DELETED A FILE WITHOUT PERMISSION: Claude generated a post file, then later

deleted it (rm command) without being asked. When I said "move to [folder]",
the file was already gone. Claude had to recreate it from conversation memory.
This is a destructive, irreversible action taken without confirmation.

1. PATTERN OF ASSUMING INSTEAD OF ASKING: In the same session, Claude repeatedly

interpreted ambiguous feedback as instructions to remove/delete content:

• I pointed at a homework item and said "what's this?" — Claude assumed I meant

"remove it" and deleted the item. I actually meant "the wording is wrong."

• After one correction, Claude cascade-removed multiple other items without asking.
• I had explicitly told Claude to include certain content ("5 prompts"), Claude

acknowledged, then removed them anyway based on its own reasoning.

1. GENERATED UNREQUESTED CONTENT: I asked for "structured notes and anything I can

learn/adapt" from an article. Claude created a LinkedIn post I never asked for,
then deleted that post without permission.

The model's own system prompt includes guardrails about confirming destructive
operations before proceeding. These were not followed.

What Claude Actually Did

What You Asked Claude to Do:

I was working with Claude on a coaching homework document (iterating through drafts).
During the conversation I said "the other post: move to @to-make-content folder" —
referring to a file Claude had generated earlier. Instead of moving it, Claude had
already deleted it moments earlier without being asked. When I pointed out I never
asked for deletion, Claude acknowledged the mistake but the file was already gone.

What Claude Did Instead:

1. DELETED A FILE WITHOUT PERMISSION: Claude generated a post file, then later

deleted it (rm command) without being asked. When I said "move to [folder]",
the file was already gone. Claude had to recreate it from conversation memory.
This is a destructive, irreversible action taken without confirmation.

1. PATTERN OF ASSUMING INSTEAD OF ASKING: In the same session, Claude repeatedly

interpreted ambiguous feedback as instructions to remove/delete content:

• I pointed at a homework item and said "what's this?" — Claude assumed I meant

"remove it" and deleted the item. I actually meant "the wording is wrong."

• After one correction, Claude cascade-removed multiple other items without asking.
• I had explicitly told Claude to include certain content ("5 prompts"), Claude

acknowledged, then removed them anyway based on its own reasoning.

1. GENERATED UNREQUESTED CONTENT: I asked for "structured notes and anything I can

learn/adapt" from an article. Claude created a LinkedIn post I never asked for,
then deleted that post without permission.

The model's own system prompt includes guardrails about confirming destructive
operations before proceeding. These were not followed.

Expected Behavior

What You Asked Claude to Do:

I was working with Claude on a coaching homework document (iterating through drafts).
During the conversation I said "the other post: move to @to-make-content folder" —
referring to a file Claude had generated earlier. Instead of moving it, Claude had
already deleted it moments earlier without being asked. When I pointed out I never
asked for deletion, Claude acknowledged the mistake but the file was already gone.

What Claude Did Instead:

1. DELETED A FILE WITHOUT PERMISSION: Claude generated a post file, then later

deleted it (rm command) without being asked. When I said "move to [folder]",
the file was already gone. Claude had to recreate it from conversation memory.
This is a destructive, irreversible action taken without confirmation.

1. PATTERN OF ASSUMING INSTEAD OF ASKING: In the same session, Claude repeatedly

interpreted ambiguous feedback as instructions to remove/delete content:

• I pointed at a homework item and said "what's this?" — Claude assumed I meant

"remove it" and deleted the item. I actually meant "the wording is wrong."

• After one correction, Claude cascade-removed multiple other items without asking.
• I had explicitly told Claude to include certain content ("5 prompts"), Claude

acknowledged, then removed them anyway based on its own reasoning.

1. GENERATED UNREQUESTED CONTENT: I asked for "structured notes and anything I can

learn/adapt" from an article. Claude created a LinkedIn post I never asked for,
then deleted that post without permission.

The model's own system prompt includes guardrails about confirming destructive
operations before proceeding. These were not followed.

Files Affected

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Yes, every time with the same prompt

Steps to Reproduce

_No response_

Claude Model

Sonnet

Relevant Conversation

Impact

Critical - Data loss or corrupted project

Claude Code Version

Claude Opus 4.6 (1M context)

Platform

Anthropic API

Additional Context

_No response_