A flaw in a common AI agent connector let attackers run commands
The April 2026 disclosure affected the protocol many AI tools use to reach files and apps, and produced more than 10 critical vulnerability reports.
Kivorent News desk, published , 1 min read
10+critical vulnerabilities filed
This is a look back at an event from . We wrote it on November 11, 2026, from the sources listed at the end.
The security firm Ox Security disclosed a flaw on April 15, 2026, in the way the Model Context Protocol connects AI tools to local programs. It allowed arbitrary operating system commands to run before the connection's handshake completed. More than 10 critical vulnerabilities were filed as a result, according to AI Weekly.
Anthropic, which created the protocol, said that cleaning up input is the responsibility of the developer building on it.
What MCP is
It is a standard that lets an AI agent use outside tools: read a file, query a database, call an app.
Why it mattered
A connector that many products share is a single point of failure. The disagreement over who should fix it, the protocol or the tools built on it, is a common one in software, and it usually ends with users exposed in the meantime.
Sources
- Generative AI: Anthropic just retook the frontier, quietly, AI Weekly, week of April 11 to 18, 2026
Internet, outages and securityAI agentsAnthropic
Spotted an error? Tell us and we will fix it.