Trajectory Labs ran 720 attacks on Claude, zero landed

Trajectory Labs threw 720 indirect prompt injection attacks at Claude's auto mode, and none broke through, per TechCrunch.

Anthropic is using the results to justify switching Claude Code's auto mode on by default, the outlet reported. That means people who never touched that setting will soon have Claude acting on its own without asking first.

The independent evaluator ran 72 attack scenarios, each ten times, against Fable 5, Opus 5, and Sonnet 5 as of July 17. Anthropic didn't design or see the scenarios in advance, which is a meaningfully different claim than a vendor grading its own homework.

That's real. A held-out third-party eval beats internal red-teaming. A clean sweep across three models and 720 attempts is a specific, checkable number, not a vibe, and I'll take that seriously. But 72 scenarios is a slice of the ways someone can hide instructions in a file, a page, or a tool result Claude reads mid-task. I don't think zero successes against a known test set tells you much about attacks nobody thought to write yet.

Watch whether Anthropic or Trajectory Labs publishes the scenario list, or repeats the test after auto mode ships to everyone. A single clean run before launch is a snapshot, not a track record.

Anthropic released the eval alongside the news that auto mode is becoming Claude Code's default, per TechCrunch.