What's this about?
Anthropic's Boris Cherny praised OpenAI for improving resistance to prompt injection, then said Claude had solved the problem “in practice” two months earlier. Meta's detailed Muse launch the same day treats prompt injection as persistent and builds multiple independent containment layers around it.
What exactly is “prompt injection”?
Prompt injection is malicious text in a webpage, email or file that tries to make an agent follow the attacker's instructions instead of the user's. It becomes dangerous when the same agent can read secrets and send data out.
What's the community saying?
Cherny's post had about 3,000 likes and 435 replies.
- Supporters like public cross-lab comparisons because they pressure competitors to improve safety.
- Security practitioners object to “solved” language when attacks, tools and deployment contexts change continually.
- Muse's architecture suggests model resistance alone is insufficient: credentials, permissions and egress still need deterministic barriers.