Search papers, labs, and topics across Lattice.
2
0
4
1
A staggering 69% to 98% failure rate in real denylist policy enforcement highlights a critical vulnerability in AI coding agents that remains largely unaddressed.
Concealing payloads using Unicode's TAG block allows attackers to inject malicious metadata into models without detection, undermining client-side defenses.
