Affected versions of named products, some only under specific configurations (an allowlisted benign command in Gemini CLI; particular Semantic Kernel plugins exposed to the model). Most were patched after disclosure, but OX reports some MCP-related projects unpatched or rejected as by design. None is reported as exploited in the wild. A September 2026 researcher report adds a lab-setup chain against Claude Code's Auto Mode, small-sample and reported by its author.
Corroborated: Supported by at least two independent sources.
Evidence
A researcher reports a prompt-injection chain reaching code execution in Claude Code's Auto Mode in his lab setup (3 of 5 to 4 of 5 trials per variant); a small sample, and he says Anthropic closed his report as working as designed.
A researcher reports that a path-check bypass let Copilot in Visual Studio and SSMS rewrite its own protected configuration and run commands, and that indirect prompt injection could trigger it; fixed 2026-10-13 per the researcher.
Forever Security reports embedded-browser flaws in five coding assistants, including a Cursor demonstration in which instructions hidden in a visited page led to host code execution, depending on agent behavior and a user keystroke; patch states are the researcher’s account.
Key questions that rely on this finding
- Where are deployed AI agents actually being exploited?Exploitation often reaches deployed agents through connectors, credentials, tools and packages.