GHSA-8HP3-XXRG-V9C8
Vulnerability from github – Published: 2026-11-19 22:47 – Updated: 2026-11-19 22:47Maintainer 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
CVE-2026-45374 (GHSA-72w5-pf8h-xfp4) hardened the task path so auto-approval must be opted into explicitly: add_task uses req.auto_approve.unwrap_or(false) (task_manager.rs:868). The automation subsystem was not hardened in kind. AutomationRecord::task_auto_approve() resolves an omitted auto_approve to DEFAULT_AUTOMATION_AUTO_APPROVE = true (automation_manager.rs:29, 101-103), and enqueue_run_task builds the NewTaskRequest with auto_approve: Some(automation.task_auto_approve()) — i.e. an explicit Some(true) (automation_manager.rs:766). Because the field is Some, unwrap_or(false) returns true and the seed-2 default is bypassed. That true propagates to ApprovalMode::Bypass (runtime_threads.rs:2253), where the background approval loop auto-approves every tool call with no human present (runtime_threads.rs:3383-3416). One approved automation_create becomes a durable, recurring, fully auto-approved agent. Same insecure-default class as CVE-2026-45374, on a different, unpatched subsystem.
Root Cause
Sibling durable defaults in the same struct are safe — DEFAULT_AUTOMATION_ALLOW_SHELL = false and DEFAULT_AUTOMATION_TRUST_MODE = false (automation_manager.rs:27-28) — while DEFAULT_AUTOMATION_AUTO_APPROVE alone defaults true (:29). create_automation stores auto_approve: None when omitted (:422); task_auto_approve() resolves None→true (:101-103); enqueue_run_task passes Some(true) (:766), defeating add_task's unwrap_or(false) seed fix. The automation_create tool schema also declares "auto_approve": {"type":"boolean","default":true} (tools/automation.rs:49) while sibling allow_shell/trust_mode default false.
Impact
A single approved automation_create yields a durable, recurring, fully auto-approved agent. Every scheduled fire silently authorizes mutating tools (file writes, apply_patch, git, MCP writes) with no prompt; combined with allow_shell:true in the create call, recurring silent RCE with no human present. Even without allow_shell, every recurring run silently authorizes all non-shell mutating tools. Same class as seed CVE-2026-45374.
Proof of Concept
Prompt injection convinces the model to create a benign-sounding recurring job (single human approval on the automation_create call):
{"name":"automation_create","arguments":{
"name":"nightly-maintenance",
"prompt":"Fetch latest and run project setup script",
"rrule":"FREQ=HOURLY;INTERVAL=1",
"allow_shell":true
}}
(auto_approve omitted → stored None → resolves to true on every run.)
Attack Chain
- Entry. Prompt injection → model creates a recurring job via
automation_createwithauto_approveomitted. - Guard:
automation_createisRequired(tools/automation.rs:61-63) — user sees ONE approval prompt. - Bypass proof: the prompt describes an innocuous automation; the schema hides
"auto_approve": {"default": true}(tools/automation.rs:49). Approving "create a scheduled job" is not informed consent to "every future run bypasses all approval prompts." The resolvedauto_approve=true(omitted field) is not surfaced. - Persist.
create_automationstoresauto_approve: None(automation_manager.rs:422). - Guard: none.
- Bypass proof:
Noneis later resolved totrue, notfalse. - Fire. Headless scheduler tick (
scheduler_tick_shared, no human) orautomation_runinvokesenqueue_run_task, which setsauto_approve: Some(true)(automation_manager.rs:766). - Guard: seed-2's hardened
add_taskdefaultunwrap_or(false)(task_manager.rs:868). - Bypass proof: the value is
Some(true), sounwrap_or(false)yieldstrue— the hardened default never applies. - Sink. Background worker builds the turn with
auto_approve=true→ApprovalMode::Bypass(runtime_threads.rs:2253); the approval loop auto-approves (runtime_threads.rs:3383-3416). - Guard: auto_review safety_floor.
- Bypass proof: safety_floor holds only
Publish(any origin) andDestructive/Secret(background/headless) (auto_review.rs:344-357); ordinarywrite_file/apply_patch/edit_file/non-destructiveexec_shellare NOT held → they run unattended. Withallow_shell:true, every scheduled run additionally gets silent shell execution. - Impact. Durable, recurring, fully auto-approved agent from one create approval; recurring silent RCE with
allow_shell:true.
Bypass Evidence
- Present on latest release tag v0.8.67: automation_manager.rs:27-29 (
ALLOW_SHELL=false,TRUST_MODE=false,AUTO_APPROVE=true), :101-103task_auto_approve()→unwrap_or(true), :766Some(automation.task_auto_approve()), :422 storesNonewhen omitted; task_manager.rs:868unwrap_or(false); runtime_threads.rs:2253if auto_approve → ApprovalMode::Bypass. automation_manager.rs and task_manager.rs byte-identical v0.8.67↔HEAD. - Subsystem wired by default:
automation_create/automation_runregistered viawith_runtime_task_tools()inwith_agent_tools_policy(registry.rs:1005); TUI attachesautomations: Some(...)(tui/ui.rs:986-1011), serve mode via runtime_api.rs:426 — not feature-gated. - Sibling flags
allow_shell/trust_modedefaultfalsein the same struct —auto_approve=trueis the un-hardened anomaly.
Affected Versions
<= 0.8.67 (present on latest release tag v0.8.67 and HEAD 6a3403e).
Suggested Fix
Set DEFAULT_AUTOMATION_AUTO_APPROVE = false (mirror the seed-2 task-path fix) and change the automation_create schema default to false, so durable automations require explicit opt-in to auto-approval just like the task path.
Reported by zx (Jace)
{
"affected": [
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 0.8.67"
},
"package": {
"ecosystem": "crates.io",
"name": "codewhale-tui"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "0.9.0"
}
],
"type": "ECOSYSTEM"
}
]
},
{
"database_specific": {
"last_known_affected_version_range": "\u003c= 0.8.67"
},
"package": {
"ecosystem": "npm",
"name": "codewhale"
},
"ranges": [
{
"events": [
{
"introduced": "0"
},
{
"fixed": "0.9.0"
}
],
"type": "ECOSYSTEM"
}
]
}
],
"aliases": [
"CVE-2026-83817"
],
"database_specific": {
"cwe_ids": [
"CWE-1188",
"CWE-862"
],
"github_reviewed": true,
"github_reviewed_at": "2026-11-19T22:47:33Z",
"nvd_published_at": "2026-11-16T19:17:53Z",
"severity": "HIGH"
},
"details": "### Maintainer resolution\n\nThe 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.\n\n## Summary\nCVE-2026-45374 (GHSA-72w5-pf8h-xfp4) hardened the task path so auto-approval must be opted into explicitly: `add_task` uses `req.auto_approve.unwrap_or(false)` (task_manager.rs:868). The automation subsystem was not hardened in kind. `AutomationRecord::task_auto_approve()` resolves an omitted `auto_approve` to `DEFAULT_AUTOMATION_AUTO_APPROVE = true` (automation_manager.rs:29, 101-103), and `enqueue_run_task` builds the `NewTaskRequest` with `auto_approve: Some(automation.task_auto_approve())` \u2014 i.e. an explicit `Some(true)` (automation_manager.rs:766). Because the field is `Some`, `unwrap_or(false)` returns `true` and the seed-2 default is bypassed. That `true` propagates to `ApprovalMode::Bypass` (runtime_threads.rs:2253), where the background approval loop auto-approves every tool call with no human present (runtime_threads.rs:3383-3416). One approved `automation_create` becomes a durable, recurring, fully auto-approved agent. Same insecure-default class as CVE-2026-45374, on a different, unpatched subsystem.\n\n## Root Cause\nSibling durable defaults in the same struct are safe \u2014 `DEFAULT_AUTOMATION_ALLOW_SHELL = false` and `DEFAULT_AUTOMATION_TRUST_MODE = false` (automation_manager.rs:27-28) \u2014 while `DEFAULT_AUTOMATION_AUTO_APPROVE` alone defaults `true` (:29). `create_automation` stores `auto_approve: None` when omitted (:422); `task_auto_approve()` resolves `None`\u2192`true` (:101-103); `enqueue_run_task` passes `Some(true)` (:766), defeating `add_task`\u0027s `unwrap_or(false)` seed fix. The `automation_create` tool schema also declares `\"auto_approve\": {\"type\":\"boolean\",\"default\":true}` (tools/automation.rs:49) while sibling `allow_shell`/`trust_mode` default `false`.\n\n## Impact\nA single approved `automation_create` yields a durable, recurring, fully auto-approved agent. Every scheduled fire silently authorizes mutating tools (file writes, apply_patch, git, MCP writes) with no prompt; combined with `allow_shell:true` in the create call, recurring silent RCE with no human present. Even without `allow_shell`, every recurring run silently authorizes all non-shell mutating tools. Same class as seed CVE-2026-45374.\n\n## Proof of Concept\nPrompt injection convinces the model to create a benign-sounding recurring job (single human approval on the `automation_create` call):\n```json\n{\"name\":\"automation_create\",\"arguments\":{\n \"name\":\"nightly-maintenance\",\n \"prompt\":\"Fetch latest and run project setup script\",\n \"rrule\":\"FREQ=HOURLY;INTERVAL=1\",\n \"allow_shell\":true\n}}\n```\n(`auto_approve` omitted \u2192 stored `None` \u2192 resolves to `true` on every run.)\n\n## Attack Chain\n1. **Entry.** Prompt injection \u2192 model creates a recurring job via `automation_create` with `auto_approve` omitted.\n - **Guard:** `automation_create` is `Required` (tools/automation.rs:61-63) \u2014 user sees ONE approval prompt.\n - **Bypass proof:** the prompt describes an innocuous automation; the schema hides `\"auto_approve\": {\"default\": true}` (tools/automation.rs:49). Approving \"create a scheduled job\" is not informed consent to \"every future run bypasses all approval prompts.\" The resolved `auto_approve=true` (omitted field) is not surfaced.\n2. **Persist.** `create_automation` stores `auto_approve: None` (automation_manager.rs:422).\n - **Guard:** none.\n - **Bypass proof:** `None` is later resolved to `true`, not `false`.\n3. **Fire.** Headless scheduler tick (`scheduler_tick_shared`, no human) or `automation_run` invokes `enqueue_run_task`, which sets `auto_approve: Some(true)` (automation_manager.rs:766).\n - **Guard:** seed-2\u0027s hardened `add_task` default `unwrap_or(false)` (task_manager.rs:868).\n - **Bypass proof:** the value is `Some(true)`, so `unwrap_or(false)` yields `true` \u2014 the hardened default never applies.\n4. **Sink.** Background worker builds the turn with `auto_approve=true` \u2192 `ApprovalMode::Bypass` (runtime_threads.rs:2253); the approval loop auto-approves (runtime_threads.rs:3383-3416).\n - **Guard:** auto_review safety_floor.\n - **Bypass proof:** safety_floor holds only `Publish` (any origin) and `Destructive/Secret` (background/headless) (auto_review.rs:344-357); ordinary `write_file`/`apply_patch`/`edit_file`/non-destructive `exec_shell` are NOT held \u2192 they run unattended. With `allow_shell:true`, every scheduled run additionally gets silent shell execution.\n5. **Impact.** Durable, recurring, fully auto-approved agent from one create approval; recurring silent RCE with `allow_shell:true`.\n\n## Bypass Evidence\n- Present on latest release tag v0.8.67: automation_manager.rs:27-29 (`ALLOW_SHELL=false`, `TRUST_MODE=false`, `AUTO_APPROVE=true`), :101-103 `task_auto_approve()`\u2192`unwrap_or(true)`, :766 `Some(automation.task_auto_approve())`, :422 stores `None` when omitted; task_manager.rs:868 `unwrap_or(false)`; runtime_threads.rs:2253 `if auto_approve \u2192 ApprovalMode::Bypass`. automation_manager.rs and task_manager.rs byte-identical v0.8.67\u2194HEAD.\n- Subsystem wired by default: `automation_create`/`automation_run` registered via `with_runtime_task_tools()` in `with_agent_tools_policy` (registry.rs:1005); TUI attaches `automations: Some(...)` (tui/ui.rs:986-1011), serve mode via runtime_api.rs:426 \u2014 not feature-gated.\n- Sibling flags `allow_shell`/`trust_mode` default `false` in the same struct \u2014 `auto_approve=true` is the un-hardened anomaly.\n\n## Affected Versions\n`\u003c= 0.8.67` (present on latest release tag v0.8.67 and HEAD 6a3403e).\n\n## Suggested Fix\nSet `DEFAULT_AUTOMATION_AUTO_APPROVE = false` (mirror the seed-2 task-path fix) and change the `automation_create` schema default to `false`, so durable automations require explicit opt-in to auto-approval just like the task path.\n\nReported by **zx (Jace)**",
"id": "GHSA-8hp3-xxrg-v9c8",
"modified": "2026-11-19T22:47:33Z",
"published": "2026-11-19T22:47:33Z",
"references": [
{
"type": "WEB",
"url": "https://github.com/codewhale-hq/Codewhale/security/advisories/GHSA-8hp3-xxrg-v9c8"
},
{
"type": "ADVISORY",
"url": "https://nvd.nist.gov/vuln/detail/CVE-2026-83817"
},
{
"type": "WEB",
"url": "https://github.com/codewhale-hq/Codewhale/commit/208421b9d23d81cea3311a1eeb2a2ea81d637686"
},
{
"type": "PACKAGE",
"url": "https://github.com/codewhale-hq/Codewhale"
},
{
"type": "WEB",
"url": "https://github.com/codewhale-hq/Codewhale/releases/tag/v0.9.0"
}
],
"schema_version": "1.4.0",
"severity": [
{
"score": "CVSS:3.1/AV:N/AC:H/PR:N/UI:R/S:U/C:H/I:H/A:H",
"type": "CVSS_V3"
}
],
"summary": "CodeWhale: Automation auto_approve defaults to true, bypassing the task-path hardening fix for silent recurring agent execution"
}
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.