Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB
CVE-2026-109561 Overview
CVE-2026-109561 is an authorization bypass vulnerability in the goose AI agent framework, affecting versions 1.28.0 through 1.43.0. The flaw allows tools to execute without user approval while the agent operates in Manual Approval mode. Persisted readOnlyHint annotations from a prior Smart Approval session are never cleared by update_goose_mode, so prompt-injected tool calls bypass the user's selected approval policy inside PermissionInspector::inspect. The issue is tracked under CWE-863: Incorrect Authorization and was fixed in goose version 1.44.0.
Critical Impact
Prompt-injected calls to tools annotated readOnlyHint: true run without prompting the user, defeating the Manual Approval control that users rely on to gate tool execution.
Affected Products
- goose versions 1.28.0 through 1.43.0
- PermissionInspector component in crates/goose/src/permission/permission_inspector.rs
- Agent sessions that switched from Smart Approval mode to Manual Approval mode
Discovery Timeline
- 2026-10-12 - CVE-2026-109561 published to NVD
- 2026-10-12 - Last updated in NVD database
- Fixed in release - goose v1.44.0
Technical Details for CVE-2026-109561
Vulnerability Analysis
The goose agent supports multiple approval modes that govern whether a user is prompted before a tool call executes. In Smart Approval mode, the framework calls apply_tool_annotations, attaching readOnlyHint: true to tools that the agent considers safe to run without prompting. The PermissionInspector::inspect function then consults these annotations to decide whether to request user approval.
When a user switches the session to Manual Approval mode via update_goose_mode, the persisted annotations from the earlier Smart Approval session are not reset. PermissionInspector::inspect continues to honor the stale readOnlyHint: true flag and auto-approves tools that the user expected to gate manually. An attacker who achieves prompt injection can target any tool carrying that annotation and trigger execution without a prompt.
Root Cause
The vulnerability stems from state that outlives its security context. apply_tool_annotations writes approval-relevant metadata onto tool descriptors, but no code path clears that metadata when the operating mode changes. PermissionInspector::inspect treats the annotation as authoritative, producing an authorization decision that no longer matches the user's current policy selection [CWE-863].
Attack Vector
Exploitation requires a session that previously ran in Smart Approval mode and later switched to Manual Approval mode. An attacker then supplies prompt-injected content, for example through an untrusted document, web page, or tool output consumed by the agent. The injected prompt instructs the model to invoke a tool whose descriptor still carries readOnlyHint: true, and the call executes without user confirmation.
// Patch excerpt: crates/goose/src/agents/agent.rs
// fix(permissions): enforce manual approval for code mode (#10528)
let goose_mode = *self.current_goose_mode.lock().await;
- if goose_mode == GooseMode::SmartApprove {
- self.tool_inspection_manager.apply_tool_annotations(&tools);
- }
-
let tool_call_cut_off = match Config::global().get_param::<usize>("GOOSE_TOOL_CALL_CUTOFF")
{
Ok(v) => v,
Source: GitHub Commit b0f4e2a
The patch removes the per-call application of tool annotations that was leaking approval state across mode changes. A second hunk in crates/goose/src/agents/platform_extensions/code_execution.rs rewrites the ToolAnnotations for the execute_bash tool so it is no longer flagged as read-only.
Detect Identity Attacks the Moment Authentication Fails
Singularity™ Identity surfaces credential misuse, lateral movement, and privilege escalation in real time, so you can contain the breach before it spreads.
Stop Identity AttacksDetection Methods for CVE-2026-109561
Indicators of Compromise
- Tool executions in a goose session logged while the session is in Manual Approval mode without a corresponding user approval event.
- Session transcripts showing a mode switch from Smart Approval to Manual Approval followed by unprompted invocations of annotated tools.
- Tool calls originating from content sources that are typical prompt-injection vectors, such as fetched web pages, documents, or MCP tool outputs.
Detection Strategies
- Inventory running goose installations and compare versions against the fixed release 1.44.0.
- Review audit logs from PermissionInspector::inspect for decisions that bypassed user prompting on tools marked readOnlyHint: true.
- Correlate agent telemetry with user input events to identify tool calls that executed without any corresponding approval interaction.
Monitoring Recommendations
- Forward goose session logs to a centralized store and alert on tool executions that lack a matching approval record.
- Monitor for use of sensitive tools such as execute_bash, execute_typescript, and file-mutating MCP tools in Manual Approval sessions.
- Track process execution spawned by goose agents on endpoints to detect downstream activity resulting from unapproved tool calls.
How to Mitigate CVE-2026-109561
Immediate Actions Required
- Upgrade goose to version 1.44.0 or later, which contains the fix merged in pull request #10528.
- Restart any long-running goose sessions after upgrading so stale in-memory tool annotations are discarded.
- Audit historical session logs for tool executions that occurred in Manual Approval mode without a user prompt.
Patch Information
The fix is delivered in goose v1.44.0. The commit b0f4e2a0f1a5507aa0a0fbe73d0b5c77a3e86845 removes the conditional call to apply_tool_annotations in the agent request path and corrects the ToolAnnotations for the bash execution tool so it is no longer treated as read-only. See the VulnCheck Security Advisory for additional context.
Workarounds
- Start a fresh goose session in Manual Approval mode rather than switching modes mid-session, so no readOnlyHint annotations persist from Smart Approval.
- Restrict goose agents from consuming untrusted input sources that can carry prompt-injection payloads until the upgrade is applied.
- Disable or remove MCP tools that mark sensitive operations as readOnlyHint: true until versions are upgraded.
# Upgrade goose to the patched release
cargo install --locked --version 1.44.0 goose-cli
# Verify the installed version
goose --version
Disclaimer: This content was generated using AI. While we strive for accuracy, please verify critical information with official sources.
Sources: National Vulnerability Database, CISA KEV, Exploit-DB

