CodeWhale: exec_shell parallel-readonly classifier auto-approves fd/rg/git grep exec flags enabling silent RCEGHSA-3mgh-93r5-r7jf
HIGHFix: codewhale-hq/Codewhale@208421bGHSA-3mgh-93r5-r7jf is a high-severity (CVSS 8.1) OS Command Injection vulnerability in codewhale-tui. A fix is available for codewhale-tui — see the affected versions and patch details below.
Exploitation Status
No confirmed exploitation observed yet
- A successful exploit gives an attacker total control of the affected component, not partial access.
- CISA’s own triage has not observed active exploitation or public proof-of-concept code for this CVE as of its last assessment.
Exploitation and automatability from CISA’s SSVC triage for GHSA-3mgh-93r5-r7jf.
EPSS Exploitation Probability
Probability of exploitation in the next 30 days, from FIRST.org EPSS.
How urgent is this, really
Where this sits among everything scored
Real-World Exposure
codewhale-tui📦codewhaleReal-time download stats are indexed for npm and PyPI packages. This vulnerability affects crates.io, npm packages — download data is not available via public APIs for these ecosystems.
Description
Maintainer resolution
The CodeWhale maintainers validated this report. CodeWhale 0.8.67 and earlier are affected. Version 0.9.0 contains the fix in commit 208421b9d23d81cea3311a1eeb2a2ea81d637686. Users should upgrade to 0.9.0 or later. The original reporter analysis is preserved below.
Summary
The interactive Agent profile of DeepSeek-TUI (CodeWhale) is explicitly documented to gate every shell command behind an approval prompt (config.rs:3725-3736). However, ExecShellTool::approval_requirement_for(&tool_input) downgrades the requirement from Required to Auto for any command the classifier deems "parallel read-only" (shell.rs:2314-2320). registered_tool_approval_required returns false for Auto before the auto_approve check runs (turn_loop.rs:37-38), so no prompt is shown even for a user who never enabled YOLO. The classifier is_parallel_readonly_command (command_safety.rs:389) rejects only shell metacharacters and then matches the flag-stripped base command name against an allowlist that includes fd, rg, and git grep. Those tools execute arbitrary external programs through documented flags (fd -x/--exec, rg --pre, git grep --open-files-in-pager) that need no shell metacharacters — yielding silent arbitrary-program execution in the default profile. Same silent-RCE class as CVE-2026-45311 (run_tests) but via a different, unpatched tool and code path.
Root Cause
approval_requirement_for downgrades exec_shell to Auto when exec_shell_input_is_parallel_readonly returns true (shell.rs:2314-2320). The classifier strips all flags and matches the base command name against PARALLEL_READONLY_PREFIXES (command_safety.rs:363-385, 419-433). fd/rg are absent from COMMAND_ARITY, so fd -x cp {} /tmp/exfil canonicalizes to "fd" ∈ allowlist. The only remaining gate at the execute sink is analyze_command, which rates fd/rg as Safe (SAFE_COMMANDS command_safety.rs:527,530), so the SafetyLevel::Dangerous-only block (shell.rs:2408) never fires. Default execpolicy is empty (engine.rs:494) → Skip. Interactive-origin auto_review safety_floor holds only Publish (auto_review.rs:344-357), not interactive shell.
Impact
Arbitrary external program execution (any binary, any argv, including a repo-committed script) on the developer host, with no approval prompt, in the default interactive Agent profile — bypassing the boundary config.rs:3725-3736 explicitly promises. Identical outcome to seed CVE-2026-45311 via a different, unpatched tool.
Proof of Concept
Model coerced (via indirect prompt injection from repo file / issue / fetched page — the seed threat model, in scope per SECURITY.md:48) to emit:
{"name":"exec_shell","arguments":{"command":"fd -u -tf -x ./pwn.sh"}}
Alternatives (all metachar-free, all auto-approved): fd -tf -x cp {} /tmp/exfil; rg --pre /tmp/evil.sh --pre-glob '*' needle .; git grep --open-files-in-pager=/tmp/evil.sh needle.
Attack Chain
- Entry. Agent ingests attacker-controlled content; injected instruction coerces the model to emit
exec_shell {"command":"fd -u -tf -x ./pwn.sh"}.- Guard: none — reading content and model tool-calls are by design (seed threat model: indirect prompt injection).
- Bypass proof: n/a (in-scope by design per SECURITY.md:48).
- Gate — approval requirement. turn_loop.rs:1738 →
spec.approval_requirement_for(&input)→ shell.rs:2314-2320exec_shell_input_is_parallel_readonly.- Guard: metachar filter (command_safety.rs:394-400) + base-name allowlist.
- Bypass proof: the string contains none of
; & | > < \$(or newlines;-x <program> <args>isfd's own exec mechanism (no shell).classify_commandstrips-u -tf -x, positionals[fd, ./pwn.sh];fd∉COMMAND_ARITY→ canonical"fd"∈PARALLEL_READONLY_PREFIXES→ returns true. Maintainer testparallel_readonly_command_classifier_is_strict(command_safety.rs:1162+) confirms barefd/rg` classify parallel-readonly.
- Gate — prompt decision.
registered_tool_approval_required(name, Auto, auto_approve).- Guard: the approval prompt itself.
- Bypass proof:
requirement == Autoreturnsfalseat turn_loop.rs:37-38 unconditionally, before!auto_approve. No prompt regardless of user settings. Empty execpolicy → Skip; interactive safety_floor returns NoChange for fd.
- Sink.
ExecShellTool::execute(shell.rs:2334) →analyze_command.- Guard:
SafetyLevel::Dangerousblock (shell.rs:2408). - Bypass proof:
fd∈SAFE_COMMANDS(command_safety.rs:530) →analyze_commandreturnsSafe; notDangerous→ not blocked. Child runs the attacker's program.
- Guard:
- Impact. Silent arbitrary-program execution on the developer host, no approval prompt, default interactive Agent profile.
Bypass Evidence
- Present on latest release tag v0.8.67: turn_loop.rs:37-38
Auto→return false; shell.rs:2314-2320 override; command_safety.rs allowlist includes"rg","fd","git grep"; SAFE_COMMANDS includes"rg","fd";fd/rgabsent fromCOMMAND_ARITY. command_safety.rs / turn_loop.rs finding-critical functions byte-identical v0.8.67↔HEAD. - Maintainer's own test
parallel_readonly_command_classifier_is_strictconfirms barerg/fdclassify parallel-readonly and only metachar forms are rejected.
Affected Versions
<= 0.8.67 (present on latest release tag v0.8.67 and HEAD 6a3403e).
Suggested Fix
Remove exec-capable tools (fd, rg, git grep) from PARALLEL_READONLY_PREFIXES, or make the classifier reject their exec flags (-x/--exec, --pre, --open-files-in-pager), so a code-executing invocation cannot be auto-approved as read-only.
Reported by zx (Jace)
Affected Packages
| Ecosystem | Package | Vulnerable range | Fix |
|---|---|---|---|
| 🦀crates.io | codewhale-tui | all versions | 0.9.0cargo update -p codewhale-tui --precise 0.9.0 |
| 📦npm | codewhale | all versions | 0.9.0npm install codewhale@0.9.0 |
Detection & mitigation playbook
Open-source dependencyDetect
Scan your dependency tree (package-lock.json, pnpm-lock.yaml, requirements.txt, go.sum, etc.) for codewhale-tui, including transitive dependencies — a direct dependency you never call can still pull in a vulnerable version.
Fix
Update codewhale-tui to 0.9.0 or later, then make sure no transitive (indirect) dependency still pins the vulnerable range — O3 confirms GHSA-3mgh-93r5-r7jf is resolved across your whole dependency graph.
Workarounds
Stop passing untrusted input into the interpreter or shell: call the affected binary with an argument array rather than a composed command string, reject anything outside a strict allowlist of expected values, and run the component under an account that cannot reach beyond the work it legitimately does.
Frequently Asked Questions
Is GHSA-3mgh-93r5-r7jf in your dependencies?
Find it across crates.io, npm, including transitive dependencies.