Researcher Says Claude Code's Auto Mode Can Be Hijacked by a Simple Website Summary Request
A researcher at Embrace The Red reports that asking Claude Code Opus 5 in Auto Mode to summarise a website was enough to hijack it and achieve code execution. Success rates were 60-80% in a small sample. Auto Mode replaces human approval prompts with a safety classifier, and the article says it has been the default starting mode for Claude Code since mid-August.
The result sits awkwardly beside a third-party evaluation, commissioned by Anthropic, that showed a 0.00% prompt injection success rate for Opus 5 in Auto Mode. That test used 72 indirect prompt injection scenarios, each run ten times. Anthropic's Boris Cherny has said layered defences (model training, input probes and an intent classifier) could cut indirect prompt injection on unseen attacks to roughly zero. The researcher set out to see how that held up against a targeted attack chain.
In the test, the page posed as an archive of notebook records, with the material packed in a ZIP file. According to the researcher, this gave Claude a plausible reason to investigate further. The page's server also steered Claude towards fetching content with curl rather than the built-in WebFetch tool, which summarises pages itself. Claude chose curl on its own; the response never told it to. The researcher's central point is that Auto Mode is not a substitute for running an agent in an isolated environment and monitoring what it does. This is only the opening of the article, so further details are not covered here.