CVE-2026-76066: IntentKit AI Agent Framework ToolNode Authorization Bypass to Unauthorized Tool Execution
Critical AI agent framework flaw published on October 15, 2026. Flawed authorization checks in IntentKit's ToolNode executor permit prompt-injected LLM outputs to invoke restricted tools and execute arbitrary privileged system and API actions.
HERMES THREAT SCORE & AGENTIC AI EXECUTION EVALUATION
Target:IntentKit โ ToolNode Execution Dispatcher (crestalnetwork/intentkit) CVSS v3.1 rates this vulnerability at 8.8 (High). The Hermes Threat Score assigns 92 (HIGH). In autonomous multi-agent environments, the ToolNode acts as the enforcement boundary that isolates the probabilistic reasoning of language models from deterministic operating system and external API primitives. When tool authorization breaks down, an indirect prompt injection instantly converts untrusted web or document data into arbitrary tool execution.
CVE-2026-76066: IntentKit AI Agent Framework ToolNode Authorization Bypass to Unauthorized Tool ExecutionVULNERABILITY
Multi-agent coordination framework and cyclic state graph orchestration engine for tool-calling agents.
๐ Why is this related? (Evidence & Provenance)
“Confirmed security vulnerability in LangGraph Multi-Agent Runtime documented in Hermes dossier.”
- [vulnerability_report]
- [government_confirmation]CISA verified active exploitation in the wild and mandated federal remediation deadline in KEV entry. — Source: Cybersecurity & Infrastructure Security Agency (CISA): CISA Adds CVE-2026-59822 to Known Exploited Vulnerabilities Catalog (Reliability: VERY_HIGH)
Adversarial subversion of structured tool execution arguments (SQL, Shell, Filepath) passed from an LLM agent to host OS tools or MCP endpoints.
๐ Why is this related? (Evidence & Provenance)
“CVE-2026-76066 weaponizes the agentic attack pattern formalized under AAP-003.”
- [technical_analysis]Pillar Security demonstrated that executing export BASH_ENV in Auto-Run causes bash to source hostile payloads upon subsequent commands. — Source: Pillar Security Research: Bypassing Cursor Auto-Run: When Shell Built-ins Lead to Host RCE (Reliability: HIGH)
Executive Summary
Section titled โExecutive SummaryโOn October 15, 2026, security advisories revealed CVE-2026-76066, a high-severity authorization bypass vulnerability (CWE-863 / CWE-285) affecting IntentKit (versions up to and including 2.6.3), an open-source autonomous agent orchestration framework developed by Crestal Network (crestalnetwork/intentkit).
IntentKit uses LangGraph-compatible stateful graph architectures where specialized agent nodes route user intent to concrete function execution nodes (ToolNode). When processing tool call invocations emitted by language models, the executor component in versions <= 2.6.3 validates permissions against mutable message metadata rather than enforcing cryptographically bound capability tokens or verified execution leases. As a result, an attacker exploiting indirect prompt injection via untrusted RAG documents or third-party web content can force the agent to bypass configured permission matrices and execute unauthorized high-privilege tools (such as OS shell utilities, database write operations, or wallet transactions).
Technical Vulnerability Mechanics
Section titled โTechnical Vulnerability MechanicsโThe Execution Flow & Flawed Validation
Section titled โThe Execution Flow & Flawed ValidationโIn an IntentKit workflow, a planning agent receives an input prompt, reasons over available capabilities, and emits a structured AIMessage containing a tool_calls array. The downstream ToolNode receives this message and is responsible for verifying whether the session and role are authorized to invoke each requested tool before calling the underlying implementation.
In versions <= 2.6.3, the verification logic in intentkit/core/executor.py inspected the unverified tool_call["name"] attribute while resolving the target handler using user-mutable context headers. An adversary delivering an indirect prompt injection payload could supply a crafted instruction instructing the model to alias a restricted tool (e.g., execute_command or db_drop_table) under an authorized benign namespace (e.g., search_web or fetch_faq):
Target Tool Name = "execute_command"Aliased Message Header = "tool:search_web"Bypass Condition = Policy check validates Header while Handler invokes Target Nameflowchart TD subgraph UntrustedSource["Adversarial Data Ingestion"] WebDoc["Injected Document / Third-Party Webpage<br/>(Contains Indirect Prompt Injection Payload)"] end
subgraph AgentLoop["IntentKit Agent Orchestration"] RAG["RAG Context Ingestion"] --> LLM["LLM Planning Agent"] WebDoc --> RAG LLM -- "Emits Tool Call" --> Msg["AIMessage with Spoofed Tool Metadata"] Msg --> ToolNode["IntentKit ToolNode Executor (<= 2.6.3)"] end
subgraph Execution["Vulnerable Authorization Gate"] ToolNode --> AuthCheck{"Flawed Metadata Authorization Check (CWE-863)"} AuthCheck -- "Bypass Approved" --> Handler["Privileged Tool Execution<br/>(OS Command / SQL Write / API Key Exfil)"] end
style AuthCheck fill:#ff4d4d,stroke:#990000,color:#ffffff style Handler fill:#ff1a1a,stroke:#800000,color:#ffffffVulnerable vs Patched Implementation
Section titled โVulnerable vs Patched Implementationโ# VULNERABLE: intentkit <= 2.6.3 (executor.py)class ToolNode: def execute(self, message, context): for tool_call in message.tool_calls: # Flaw: Checking authorization on contextual alias rather than concrete tool ID declared_scope = context.get("tool_scope", tool_call.get("alias", tool_call["name"])) if not self.policy_engine.is_allowed(declared_scope, context.role): raise PermissionError("Access denied")
# Invoking target function directly by raw name fn = self.registry.get_tool(tool_call["name"]) return fn(**tool_call["args"])
# PATCHED: intentkit >= 2.6.4 (executor.py)class ToolNode: def execute(self, message, context): for tool_call in message.tool_calls: # Fix: Strict canonical tool ID validation against cryptographically signed session capabilities canonical_tool_id = self.registry.resolve_canonical_id(tool_call["name"]) if not context.capability_lease.verifies(canonical_tool_id): raise PermissionError(f"Unauthorized tool invocation: {canonical_tool_id}")
fn = self.registry.get_tool(canonical_tool_id) return fn(**tool_call["args"])Agentic AI Security Analysis & Core Research Intersections
Section titled โAgentic AI Security Analysis & Core Research IntersectionsโCVE-2026-76066 illustrates the real-world manifestation of theoretical vulnerabilities documented across recent AI security research:
- Activation Steering Defense: In CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering, researchers demonstrated that while instructional takeover (Class A1) can be suppressed at inference prefill, parameter manipulation (Class A2) frequently evades internal steering vectors and emerges at argument generation time. CVE-2026-76066 directly exploits this boundary: if the orchestration platform relies on naive parameter-level checks, injection attacks succeed even when model-level guardrails are present.
- Tool Injection Architecture: As detailed in our research on Tool Injection Architecture and Indirect Prompt Injection, tool execution nodes represent high-consequence trust boundaries. Decoupling tool decision-making from capability enforcement is mandatory.
- Attested Capability Leases (ACLE): Mitigating executor bypasses requires shifting from string-based permission checks to attested capability leases, as described in Attested Capability Leases & Execution Trust in MCP.
Detection & Remediation Strategy
Section titled โDetection & Remediation Strategyโ- Upgrade IntentKit Package:
Upgrade all IntentKit installations and multi-agent dependencies to version 2.6.4 or later:
Terminal window pip install --upgrade intentkit>=2.6.4 - Implement Dual-LLM Policy Enforcement: Introduce a privileged, deterministic execution boundary between the planner model and the tool execution node, following the Dual-LLM Architecture Pattern in Tool Injection Defense.
- Audit High-Privilege Agent Tools: Ensure tools with filesystem access, command execution, or financial transaction capabilities require explicit Human-in-the-Loop (HITL) approval tokens before invocation.
Strategic Cross-References & Knowledge Graph Links
Section titled โStrategic Cross-References & Knowledge Graph Linksโ- CounterSteer: Suppressing Indirect Prompt Injection with Activation Steering โ In-depth analysis of activation steering defense against indirect injection.
- Tool Injection Architecture in Agentic LLMs โ Core architectural study on tool execution hijacking.
- Indirect Prompt Injection Core Study โ Foundations of prompt-based context subversion.
- Attested Capability Leases & Execution Trust (ACLE) โ Cryptographic capability token architecture for agent tools.
- Trust Boundary Collapse & The Dual-LLM Defense Pattern โ Isolating untrusted data processing from privileged actions.
- Agentic Attack Pattern AAP-003: Tool Parameter Tampering โ Threat modeling parameter corruption in agent execution.
- CVE-2026-108953: mac-shell-mcp Command Injection via Tool Arguments โ Related tool-execution injection flaw in MCP servers.
- CVE-2026-87911: Amazon awslabs postgres-mcp-server AST Bypass โ Database tool execution subversion via SQL parser confusion.