Cursor, which was not too long ago acquired by SpaceX for $60 billion in inventory, produces one of the broadly used AI-assisted coding instruments used within the enterprise house. The 2 flaws have been patched in model 3.0 of the Cursor IDE, which was launched in April.
Native vulnerability in LLMs
Giant Language Fashions (LLMs) are natively weak to malicious directions that might be hidden contained in the content material they course of. That is significantly harmful within the age of agentic AI, the place LLMs are mixed with a wide range of instruments, together with browsers and APIs that permit them to entry a wide range of third-party public content material, resembling parsing internet pages in search outcomes and RSS feeds, code in repositories, feedback in bug trackers, emails in customers’ inboxes, and their paperwork.
Defending AI instruments from immediate injection may be very arduous, and normally includes a layered method, together with guardrails constructed into the mannequin by the AI lab that created it, directions in system prompts to deal with sure content material as passive information, supervisor fashions operating on prime of the LLMs that course of information, conventional key phrase filtering, context segmentation, granular entry controls, including people again into the loop to approve delicate operations and extra.