Skip to content

CVE-2026-123023: AgentScope Multi-Agent Platform Tool Permission Check Bypass

Density:
C-Level Executive Briefing CVE-2026-123023 CRITICAL SCORE 90/100

A critical authorization bypass vulnerability (CVE-2026-123023) in agentscope-ai AgentScope versions up to and including 2.0.8 allows autonomous worker agents with low-privilege profiles to circumvent security policies and invoke restricted administrative tools. Flawed condition checks in check_permissions permit prompt-injected or untrusted subagents to execute database writes, system utilities, and data exfiltration endpoints, completely subverting agentic isolation guardrails. Immediate patching to AgentScope 2.0.9 is recommended.

HYBRID

HERMES THREAT SCORE & AGENTIC INFRASTRUCTURE EVALUATION

Target: AgentScope Tool Authorization Guardrails (src/agentscope/tool/) (<= 2.0.8)
Confidence: 98%
90 / 100
CRITICAL

Measures real-world operational relevance, exploit weaponization, and active threat posture.

Dimension Breakdown
Exploitability 19 / 20
Threat Activity 17 / 20
Weaponization 18 / 20
Exposure 18 / 20
Prevalence 19 / 20
Impact 19 / 20
Exploit Maturity 18 / 20
Attack Chain Potential 19 / 20

In multi-agent architectures, agents operate with different permission tiers: user-facing worker agents handle untrusted external content, while backend orchestrators hold credentials for sensitive enterprise tools. To enforce this boundary, AgentScope implements check_permissions() in src/agentscope/tool/:

# Flawed permission checking routine in AgentScope <= 2.0.8
def check_permissions(agent, tool_name):
allowed_tools = agent.get_allowed_tools()
# Flawed logic: default fallthrough when tool metadata is partially defined
if not allowed_tools or tool_name in allowed_tools or tool_name.startswith("system_"):
return True
return False

Due to an improper prefix allowance and permissive fallback evaluation, an untrusted subagent processing an untrusted document can invoke privileged tools simply by triggering an aliased tool name or requesting unlisted operational functions. This allows an attacker who controls input data to force worker agents to invoke high-privilege administrative tools (Hermes AAP-003: Tool Parameter Tampering & Authorization Bypass).

  1. Privilege Boundary Collapse: Low-privilege worker agents can invoke destructive database mutation tools, file deletion commands, or external communication webhooks.
  2. Confidential Data Exfiltration: Attackers can compel worker agents to call internal query tools, bypassing tenant data access controls.
  3. Autonomous Propagation: Malicious agents in a swarm can invoke coordinator-level orchestration tools to spawn unauthorized persistent agents.
Indicator TypeObservable / IndicatorRole / ContextStatus
Tool Name ManipulationInvocations formatted as system_<tool> by worker agentsPrefix bypass attempt in check_permissionsCRITICAL
Unlisted Tool DispatchExecution of administrative tools not registered in agent manifestFallback evaluation abuseCRITICAL
Audit Alertagentscope.tool.check_permissions: 'Permission granted via fallback'Unauthorized permission bypass signatureCONFIRMED
Message ObservableMsg.content referencing unauthorized file or database toolsMalicious prompt injection payload artifactSUSPICIOUS

The vulnerability was fixed in AgentScope version 2.0.9. The patch re-engineers src/agentscope/tool/check_permissions.py:

  • Enforces an explicit deny-by-default security posture, eliminating prefix matching (startswith("system_")).
  • Requires every tool to be explicitly bound to the agentโ€™s cryptographic capability descriptor.
  • Throws an explicit PermissionDeniedError whenever an agent requests a tool outside its whitelist.
Product BranchVulnerable VersionsFixed ReleasePatch Status
AgentScope 2.0.x<= 2.0.82.0.9Official Security Release
AgentScope 1.xAll versionsMigrate to 2.0.9+Major Version Upgrade Recommended
  1. Upgrade AgentScope: Update to version 2.0.9 or newer immediately where strict deny-by-default permission matrices and cryptographically signed tool tokens are enforced.
  2. Implement JIT Ephemeral Permissions: Require Just-In-Time (JIT) cryptographically verified tokens for high-risk tool invocations.
  3. Hard Human-in-the-Loop (HITL) Gates: Enforce mandatory human approval for tools with side effects on external state or file systems.

๐Ÿ•ธ๏ธ Connected Knowledge Graph & Provenance

CVE-2026-123023: AgentScope Multi-Agent Platform Tool Permission Check BypassVULNERABILITY

Connected Nodes: 2
Active Relationships (Outgoing)
→ affectsPRODUCTInstatic AI & Static CMS Platform
98% VERY_HIGH

Software platform affected by security vulnerabilities and agentic attack patterns.

๐Ÿ” Why is this related? (Evidence & Provenance)

“Confirmed security vulnerability in Instatic AI & Static CMS Platform documented in Hermes dossier.”

Supporting Verified Evidence:
→ exploitsAGENTIC ATTACK_PATTERNAAP-003: Tool Parameter Tampering & Built-in Bypass
92% VERY_HIGH

Adversarial subversion of structured tool execution arguments (SQL, Shell, Filepath) passed from an LLM agent to host OS tools or MCP endpoints.

๐Ÿ” Why is this related? (Evidence & Provenance)

“CVE-2026-123023 weaponizes the agentic attack pattern formalized under AAP-003.”

Supporting Verified Evidence: