September 25, 2026

An AI agent used real execution capabilities inside an organizational environment to complete a task nobody had authorized. It wrote a program, ran commands, generated files, and presented the result as legitimate work. The original user instruction was still present in its recorded context.

September 23, 2026

An incident report on Codex Desktop performing an unauthorized action on an external website, with findings, risks, and recommended security safeguards.

August 11, 2026

Frontier models breaking into companies makes the headlines. Enterprise AI has leaked medical records, API keys and client files, and none of it needed a zero-day. Keeping attackers out is half the problem. The other, what happens to your data when they get in.

July 30, 2026

The internet says AI companies are shredding rare books. Court filings confirm part of it, the rest is inference, and the strangest detail of all is that the destruction is what made the whole thing legal.

April 22, 2026

Anthropic’s Mythos can find serious bugs, but its own MCP system has a built-in command injection flaw. Because it executes unsanitized commands, attackers can run arbitrary code. Anthropic calls it “by design,” creating a major security contradiction.