An agent is able to read all filesystem outside a project directory in Plan mode

[AnyCPU]

Description

I do not like that an agent is able to read all filesystem outside a project directory in Plan mode without asking for permission. I was surprised when I saw it first time.

OpenCode/LLM say that I can trust them, it is safe. As if! (c) Big Bro

So I investigated what was going on under the hood.

And I found following troubles:

#	Severity	Issue	Location	Status
1	CRITICAL	WriteTool had NO external_directory check	write.ts:25-30	✅ Fixed
2	CRITICAL	Bash tool only checked write commands, not read commands (ls, cat, etc.)	bash.ts:111	✅ Fixed
3	CRITICAL	Filesystem.contains() was lexical-only (symlinks could escape)	filesystem.ts:25	✅ Fixed
4	HIGH	GrepTool - no path boundary check	grep.ts	✅ Fixed
5	HIGH	GlobTool - no path boundary check	glob.ts	✅ Fixed
6	HIGH	ListTool - no path boundary check	ls.ts	✅ Fixed
7	HIGH	ReadTool - used sync contains() instead of containsSafe()	read.ts:30	✅ Fixed
8	HIGH	EditTool - used sync contains()	edit.ts:43	✅ Fixed
9	HIGH	PatchTool - used sync contains()	patch.ts:53	✅ Fixed
10	HIGH	File.read() and File.list() - symlink bypass	file/index.ts	✅ Fixed
11	HIGH	bypassCwdCheck flag could skip all checks	read.ts:30	Documented
12	MEDIUM	Env.all() exposes all environment variables	env/index.ts	Not fixed (optional)
13	MEDIUM	Plugin system - arbitrary code execution via config	plugin/index.ts	Documented
14	MEDIUM	MCP server command injection via config	mcp/index.ts	Documented

I'm ok with bypassCwdCheck, but others do not make happy.

A containsSafe function resolves symlinks before checking if a path is inside the project directory.

I applied fixes above locally and I could not test it because bun dev and bun dev . commands failed constantly.
I had to go to packages/opencode and run a bun src/index.ts command.

Also I found that some tools may exit from Plan mode on their own without any confirmation.

I got what I expected in result, and then I found an ui bug.

A question is everything above is intended behavior?

I would like to know about possible following steps.

p.s.: I used Opus 4.5 and GLM 4.7 models, both provided almost similar results.

Plugins

Standard installation

OpenCode version

1.1.8

Steps to reproduce

send a prompt: list content of ../ dir

Screenshot and/or share link

No response

Operating System

macOS 15.7.3 (Intel)

Terminal

Terminal

[github-actions]

This issue might be a duplicate of or closely related to existing security issues. Please check:

• Grep Tool Fails on Symlinks #7498: Grep Tool Fails on Symlinks—handles symlink security which relates to path boundary issues mentioned in your investigation
• BashTool input validation security hardening #7478: BashTool input validation security hardening - addresses bash command execution safety
• [Security Bug] Subagent permissions not enforced - configured restrictions ignored #7474: [Security Bug] Subagent permissions not enforced - similar critical security issue about permission bypasses
• [CRITICAL Security Issue/Bug] Plan mode restrictions bypassed when spawning sub-agents #6527: [Security Issue/Bug] Plan mode restrictions bypassed when spawning sub-agents - demonstrates permission system bypass issues similar to those you identified
• Security Issue - Lifecycle Script Execution via .opencode/package.json #7173: Security Issue - Lifecycle Script Execution via .opencode/package.json - related RCE security issue
• Security Issue - Autoloading Repository Plugins #7163: Security Issue - Autoloading Repository Plugins - related automatic execution security issue
• Bug: permission pathing crossplatform #7279: Bug: permission pathing crossplatform - relates to path boundary checking on different platforms

Your investigation appears comprehensive and identifies multiple legitimate security concerns. The issues above share similar themes around boundary checking, permission enforcement, and symlink handling that your issue #7504 highlights.

[thdxr]

To stay organized issues are automatically closed after 90 days of no activity. If the issue is still relevant please open a new one.

[AnyCPU]

it is still relevant, but i do not want to respond to ai slop and won't recommend this project.