#prompt-injection_1
-
Auto Mode in Claude Code Is No Cure for Prompt Injection
Researcher Johann Rehberger has demonstrated an attack that defeats Claude Code's protective mode 80% of the time — and sometimes the protection itself stops the agent from shutting down malicious code.
-
“Mind Viruses” Hop Between AI Agents Through Memory Files
Researchers at Anthropic and EPFL have shown that self-propagating payloads can flow from one autonomous AI agent to another through the editable files that store state between sessions. In the real world, the technique hasn't worked so far.
-
Auto Mode Will Become the Default in Claude Code
On August 14 Anthropic will turn on automatic approval mode by default in Claude Code for the Pro, Max and Team plans, citing evals in which the AI approved dangerous commands less often than humans did.
-
Anthropic says Opus 5 is the hardest to prompt-inject
Boris Cherny of Anthropic notes that the new Opus 5 model is extremely hard to hack with prompt injection — a result he rates above its benchmark scores.