Key Info

TypeSafe AI red-teamed Jev 1.13 using their DTap platform and uncovered serious safety gaps: 70.1% ASR under direct misuse and 43.5% ASR under indirect prompt injection.

Highlights

  • Jev 1.13 shows high vulnerability to direct misuse attacks (70.1% ASR).
  • Indirect prompt injection yields 43.5% attack success rate.
  • Findings emphasize the danger of agent platforms when misused by attackers.
  • TypeSafe AI advocates for composability but underscores the urgent need for safety improvements.