Vulnerability / research
LangAlpha research links retrieved web instructions to sandbox code execution
Reported
Published here
What happened
Lasso demonstrated retrieved web instructions triggering a shell tool without approval in a local LangAlpha harness. The reported design could expose secrets available inside a user's workspace sandbox.
Parties and reporting
AI tools and systems: LangAlpha, Bash and ExecuteCode agent tools, Claude Sonnet 4.5 in the research harness
Outcome
CVE-2026-109440 covers versions through 2026.09.07. The reviewed sources identify no fixed release or malicious exploitation.
Operating lesson
Require approval for sensitive code execution, isolate secrets behind mediated tools, and constrain sandbox network access.
Evidence and limits
Date follows the CNA's public-disclosure field. Lasso's advisory, linked under parties, is undated. Its benign-command demonstration mocked Daytona locally and did not contact the live SaaS; production secret theft and sandbox escape were not demonstrated.
Sources
- VulnCheck CNA record, including October 15 public-disclosure date and researcher reference
Primary source ·
- VulnCheck LangAlpha advisory
Primary source ·