Track papers across labs, topics, and trends.
Search papers, labs, and topics across Lattice.
Affiliation:
2
0
2
0
Indirect prompt injections can exploit framing strategies to bypass defenses, turning a 0% success rate into 100% with simple rewording techniques.
Forged-reasoning attacks can completely bypass traditional defenses, but Proof-of-Execution Memory ensures that only verified actions are accepted, achieving zero successful attacks across tested models.