CVE-2026-58481

Network-AI from 5.14.0 before 5.15.5 contains a missing authentication vulnerability in the A2AServer class that allows unauthenticated attackers to access task operations when no secret is configured. Attackers can send JSON-RPC requests to tasks/send, tasks/get, and tasks/cancel without a Bearer token to execute tasks, read results, and cancel tasks.

Network-AI before 5.15.5 contains a path traversal vulnerability in InputSanitizer.sanitizePath() that allows attackers to escape the base directory using sibling-directory paths. Attackers can supply paths whose directory name shares a string prefix with the base, bypassing the unanchored startsWith check to read, list, or write files outside the sandbox.

Network-AI from 5.13.4 before 5.15.5 contains a path traversal vulnerability in AgentVCR that allows attackers controlling the cassette name to read or write arbitrary .json files. Attackers can pass traversal sequences to load() or save() to disclose sensitive JSON data or overwrite policy files such as AuthGuardian trust-level configuration.

Network-AI from 5.13.4 before 5.15.5 contains a path traversal vulnerability in FileJobStore that allows attackers to read or delete arbitrary .json files by supplying traversal sequences in the job id. Attackers controlling the id passed to FileJobStore.get(), FileJobStore.delete(), or JobQueue.getJob() can disclose credentials and grant tokens or delete configuration files to weaken AuthGuardian authorization.

Vulnerability in Network-AI (TypeScript/Node.js) before version 5.9.1 allows bypassing the shell command allowlist. An attacker can execute arbitrary commands by injecting special characters (e.g., `;`).

Network-AI (npm network-ai) before 5.15.2 contains an origin validation error in McpSseServer._handleRequest in lib/mcp-transport-sse.ts that reflects localhost Origins into Access-Control-Allow-Origin regardless of bind interface. When the server binds a non-loopback address, an attacker page with a localhost Origin visited by a victim can read unauthenticated endpoints like /health cross-origin.

Network-AI from 5.14.0 before 5.15.2 contains a security control bypass vulnerability in the ClaudeHookBridge PreToolUse hook, which matches deny patterns only against the first present candidate field. A prompt-injected agent can place dangerous payloads in fields like Write content or an MCP script parameter beside a benign file_path or command, evading denyPatterns to write web shells or execute scripts.

Network-AI 5.0.0 before 5.15.2 contains a cross-site WebSocket hijacking vulnerability in DashboardServer that allows remote attackers to read live agent topology because handleUpgrade performs no Origin or token validation. Attackers can lure a victim running the dashboard to a malicious page that opens a WebSocket to the loopback dashboard and exfiltrates agent IDs, roles, task descriptions, and token budgets.

Network-AI before version 5.15.1 contains a security matcher bypass vulnerability where SandboxPolicy evaluates raw command strings with quotes preserved while the executor tokenizes commands by stripping quotes before execution. Attackers can craft quoted commands that evade blocklist checks and approval gates while the executor runs the identical unquoted dangerous argv.

Network-AI ClaudeHookBridge before 5.15.1 truncates the target string to 500 characters before evaluating denyPatterns, while Claude Code executes the full untruncated command. Attackers can position dangerous content past byte 500 in a Bash command field to bypass the operator's hard-deny list and execute arbitrary commands.