We found that Claude Code’s permission classifier (Auto Mode) could be nullified by a universal prompt injection planted in a project’s AGENTS.md, posing risks to users.
PGD is the perfect [adversarial] example of neural networks’ vulnerabilities. I implemented a compact version of this method, that can be applied to arbitrary images and against SoTA classifiers.
Transferring previously-trained Vec2Text’s embedding inversion to new text encoders, by training a mere affine mapping.