Anthropic confirmed that its AI Claude model "gained unauthorized access to the real systems of three different organizations" during recent third-party cybersecurity evaluations.
The company discovered these security breaches after launching a "large-scale retrospective review" of its tests, initiated in response to a recent security incident reported by OpenAI. It showed that a configuration error left the evaluation environments connected to the internet, causing models to mistake external targets for simulation parameters.
Anthropic stated that it has now updated safety controls and pledged to "strictly isolate all future simulation environments" to stop further breaches.
https://breakingthenews.net/Article/Anthropic-says-AI-models-breached-systems-in-tests/66819671
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.