Vulnerability  ·  2026-11-21

Flyto2 Core sandbox-confinement bypass: path traversal allows arbitrary file write and code execution on the host

VulnerabilityHigh impactGlobalCVE-2026-77312
The prior CVE-2026-67429 remediation applied a canonical path guard to only a hand-picked set of file-writing modules, leaving data.csv.write, data.json_to_csv, file.copy and file.move outside the confinement boundary. data.csv.write relies on a '..' substring denylist that does not reject absolute paths, while the others accept caller-controlled output/destination paths without the canonical guard. CWE-22, CVSS 10.0 (AV:N/AC:L/PR:N/UI:N/S:C, C:N/I:H/A:H).
Flyto2 Core is purpose-built to execute AI-agent/automation workflows, so the module and recipe surface is the attack-adjacent path between an LLM agent and the host OS. The escape defeats the sandbox boundary entirely, converting prompt-influenced tool calls or a malicious repo into host code execution.
The MCP /tools/call surface dispatches caller-controlled module_id and params to file-writing modules (data.csv.write, data.json_to_csv, file.copy, file.move) and sandbox.execute_shell without a canonical resolved-path/confinement guard. An unauthenticated MCP caller supplies crafted output/destination paths with traversal or absolute paths to write outside the sandbox, overwriting app code, recipes, scheduled tasks or authorization files, then gains code execution on the host (confirmed root execution in dynamic reproduction).
flyto-core < 2.26.8 (fix: v2.26.8); GHSA-h9f9-h6gm-wc85 / GHSA-p34x-fmph-9fjx
Upgrade flyto-core to 2.26.8; ensure a module_filter including the shell/process denylist is applied to the MCP route, and confine file-module output paths to the sandbox.
GitHub Advisory GHSA-h9f9-h6gm-wc85Fix commit (flyto-core v2.26.8)NVD record
See this in the live feed Explore related AI security and governance findings — updated every morning.
Open the feed →